model_id stringclasses 11
values | model_display_name stringclasses 11
values | publisher stringclasses 8
values | quantization stringclasses 3
values | format stringclasses 1
value | regime stringclasses 5
values | hardware dict | runtime dict | benchmark dict | evaluation dict | session dict |
|---|---|---|---|---|---|---|---|---|---|---|
Qwen/Qwen3.6-35B-A3B | qwen3.6-35b-a3b | qwen | Q4_K_S | GGUF | moe_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "Windows 11"
} | {
"engine": "LM Studio",
"context_length": 16384,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": true,
"temperature": 0.7,
"top_p": 0.9,
"ncmoe": null,
"ngl": null
} | {
"tok_per_sec": 7.4,
"ttft_seconds": 1.75,
"tokens_generated": 761,
"vram_used_gb": null,
"ram_used_gb_peak": 31,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": 4,
"hit_length_target": true,
"notes": "RAM-ceilinged on 32GB system. Reproduces 5.5x slower than the public X-post claim of 41 tok/sec — same author's thread requires 64GB RAM as the unstated prerequisite. GGUF repackaged by bartowski."
} | {
"date": "2026-05-06T00:00:00",
"run_type": "warm"
} |
Qwen/Qwen3.6-35B-A3B | qwen3.6-35b-a3b | qwen | Q4_K_S | GGUF | moe_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "Windows 11"
} | {
"engine": "LM Studio",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": true,
"temperature": 0.7,
"top_p": 0.9,
"ncmoe": null,
"ngl": null
} | {
"tok_per_sec": 7.43,
"ttft_seconds": 1.85,
"tokens_generated": 796,
"vram_used_gb": null,
"ram_used_gb_peak": 31,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": 4,
"hit_length_target": true,
"notes": "Doubling context from 16K to 32K barely moved tok/sec (7.40 → 7.43). Flat curve = constant overhead dominates, not KV-attention-scan cost. Consistent with disk-paging bottleneck rather than RAM-bandwidth bottleneck. GGUF repackaged by bartowski."
} | {
"date": "2026-05-06T00:00:00",
"run_type": "warm"
} |
Qwen/Qwen3.6-35B-A3B | qwen3.6-35b-a3b | qwen | IQ4_XS | GGUF | moe_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "Windows 11"
} | {
"engine": "LM Studio",
"context_length": 16384,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": true,
"temperature": 0.7,
"top_p": 0.9,
"ncmoe": null,
"ngl": null
} | {
"tok_per_sec": 7.4,
"ttft_seconds": null,
"tokens_generated": null,
"vram_used_gb": null,
"ram_used_gb_peak": 31,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "Speed-only run, no quality eval captured. Smaller quant (~2GB less than Q4_K_S, 18.81GB vs 20.59GB) tested to isolate quant size as variable. Same tok/sec confirms quant size is not the lever — RAM ceiling is the binding constraint regardless of weight f... | {
"date": "2026-05-06T00:00:00",
"run_type": "warm"
} |
Qwen/Qwen3.6-35B-A3B | qwen3.6-35b-a3b | qwen | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (llama.cpp b9049-2496f9c14)",
"context_length": 16384,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 29.7,
"ttft_seconds": 2.44,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=32 (8 layers experts on GPU, 32 on CPU). ngl=999, threads=6. Prompt: 20.46 tok/s. Total: 36.9s. VRAM ~6649 MiB / 8187 MiB. CPU-mapped experts: 16581 MiB. Quantized by Unsloth (UD). Baseline partial offload — clean, no VRAM pressure. 4x faster than ... | {
"date": "2026-05-07T00:00:00",
"run_type": "cold_server"
} |
Qwen/Qwen3.6-35B-A3B | qwen3.6-35b-a3b | qwen | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (llama.cpp b9049-2496f9c14)",
"context_length": 16384,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 32,
"ttft_seconds": 1.18,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=30 (10 layers experts on GPU, 30 on CPU). ngl=999, threads=6. Prompt: 42.48 tok/s. Total: 33.2s. Sweet spot for 16K context. 7.7% faster decode than ncmoe 32. 4.3x faster than full offload."
} | {
"date": "2026-05-07T00:00:00",
"run_type": "cold_server"
} |
Qwen/Qwen3.6-35B-A3B | qwen3.6-35b-a3b | qwen | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (llama.cpp b9049-2496f9c14)",
"context_length": 16384,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 16.33,
"ttft_seconds": 2.05,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=28 (12 layers experts on GPU, 28 on CPU). ngl=999, threads=6. Prompt: 24.38 tok/s. Total: 64.7s. VRAM pressure triggered — CUDA page faults through PCIe halved throughput. Sharp cliff, not gradual."
} | {
"date": "2026-05-07T00:00:00",
"run_type": "cold_server"
} |
Qwen/Qwen3.6-35B-A3B | qwen3.6-35b-a3b | qwen | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (llama.cpp b9049-2496f9c14)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 35.36,
"ttft_seconds": 1.44,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=30 (10 layers experts on GPU, 30 on CPU). ngl=999, threads=6. Prompt: 34.72 tok/s. Total: 30.4s. Best overall config. Hybrid SSM+attention means KV cache only scales for 10/40 attention layers — 32K is nearly free. 4.8x faster than full offload. Re... | {
"date": "2026-05-07T00:00:00",
"run_type": "cold_server"
} |
Qwen/Qwen3.6-35B-A3B | qwen3.6-35b-a3b | qwen | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (llama.cpp b9049-2496f9c14)",
"context_length": 65536,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 17.41,
"ttft_seconds": 2.83,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=30 (10 layers experts on GPU, 30 on CPU). ngl=999, threads=6. Prompt: 17.70 tok/s. Total: 61.7s. VRAM cliff at 65K — KV cache for 10 attention layers at q8_0 grows to ~680 MiB, pushing total GPU past ~7 GB. Same ~50% throughput drop as ncmoe 28 at ... | {
"date": "2026-05-07T00:00:00",
"run_type": "cold_server"
} |
google/gemma-4-26B-A4B | gemma-4-26b-a4b | google | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 16384,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 27.92,
"ttft_seconds": 1.45,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=24 (6 layers experts on GPU, 24 on CPU). ngl=999, np=1, threads=6. Prompt: 43.54 tok/s. Total: 38.1s. VRAM projected 6119 MiB / 7063 MiB free (944 MiB headroom). 128 experts/layer, 8+1 shared active. Comfortable headroom but fewer expert layers on ... | {
"date": "2026-05-13T00:00:00",
"run_type": "cold_server"
} |
google/gemma-4-26B-A4B | gemma-4-26b-a4b | google | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 16384,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 29.19,
"ttft_seconds": 1.59,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=23 (7 layers experts on GPU, 23 on CPU). ngl=999, np=1, threads=6. Prompt: 39.67 tok/s. Total: 36.7s. VRAM projected 6573 MiB / 7063 MiB free (490 MiB headroom). SWEET SPOT at 16K. Best decode speed with safe VRAM margin."
} | {
"date": "2026-05-13T00:00:00",
"run_type": "cold_server"
} |
google/gemma-4-26B-A4B | gemma-4-26b-a4b | google | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 16384,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 28.34,
"ttft_seconds": 1.57,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=22 (8 layers experts on GPU, 22 on CPU). ngl=999, np=1, threads=6. Prompt: 40.08 tok/s. Total: 37.7s. VRAM projected 7026 MiB / 7063 MiB free (37 MiB headroom). Razor-thin margin — works but risky for sustained use."
} | {
"date": "2026-05-13T00:00:00",
"run_type": "cold_server"
} |
google/gemma-4-26B-A4B | gemma-4-26b-a4b | google | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 16384,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 24.5,
"ttft_seconds": 1.82,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=20 (10 layers experts on GPU, 20 on CPU). ngl=999, np=1, threads=6. Prompt: 34.54 tok/s. Total: 43.6s. VRAM projected 7934 MiB / 7063 MiB free (-871 MiB over budget). CUDA page faults, 16% speed drop. Gentler cliff than Qwen (-50% at same over-budg... | {
"date": "2026-05-13T00:00:00",
"run_type": "cold_server"
} |
google/gemma-4-26B-A4B | gemma-4-26b-a4b | google | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 29.27,
"ttft_seconds": 1.14,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=23 (7 layers experts on GPU, 23 on CPU). ngl=999, np=1, threads=6. Prompt: 55.31 tok/s. Total: 36.1s. VRAM projected 6743 MiB / 7063 MiB free (320 MiB headroom). BEST OVERALL CONFIG. 32K context nearly free thanks to 5:1 SWA-to-full attention patte... | {
"date": "2026-05-13T00:00:00",
"run_type": "cold_server"
} |
google/gemma-4-26B-A4B | gemma-4-26b-a4b | google | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 65536,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 25.78,
"ttft_seconds": 1.16,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=23 (7 layers experts on GPU, 23 on CPU). ngl=999, np=1, threads=6. Prompt: 54.12 tok/s. Total: 40.9s. VRAM projected 7083 MiB / 7063 MiB free (-20 MiB over budget). Gentle degradation (-12% vs 16K). SWA pattern keeps KV at 839 MiB. No catastrophic ... | {
"date": "2026-05-13T00:00:00",
"run_type": "cold_server"
} |
google/gemma-4-E2B-it | gemma-4-e2b | google | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 65536,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 117.8,
"ttft_seconds": null,
"tokens_generated": 512,
"vram_used_gb": 2.53,
"ram_used_gb_peak": null,
"prompt_id": "tcp-udp-explanation"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "Baseline at low context depth (44 prompt tokens). ngl=999, np=1, threads=6. Full GPU offload, 2593 MiB VRAM used, 5595 MiB free. Dense model, no -ncmoe needed. Built-in thinking mode (reasoning_content). PLE not fully implemented in llama.cpp."
} | {
"date": "2026-05-14T00:00:00",
"run_type": "cold_server"
} |
google/gemma-4-E2B-it | gemma-4-e2b | google | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 65536,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 113.1,
"ttft_seconds": null,
"tokens_generated": 512,
"vram_used_gb": 2.54,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-8k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~8K context fill (6280 prompt tokens). Prompt eval 7761 tok/s. VRAM 2605 MiB. Minimal degradation from baseline."
} | {
"date": "2026-05-14T00:00:00",
"run_type": "warm_server"
} |
google/gemma-4-E2B-it | gemma-4-e2b | google | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 65536,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 107.3,
"ttft_seconds": null,
"tokens_generated": 512,
"vram_used_gb": 2.55,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-16k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~16K context fill (12530 prompt tokens). Prompt eval 7317 tok/s. VRAM 2617 MiB. -9% from baseline."
} | {
"date": "2026-05-14T00:00:00",
"run_type": "warm_server"
} |
google/gemma-4-E2B-it | gemma-4-e2b | google | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 65536,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 87.7,
"ttft_seconds": null,
"tokens_generated": 512,
"vram_used_gb": 2.58,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-32k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~32K context fill (25030 prompt tokens). Prompt eval 6111 tok/s. VRAM 2641 MiB. -26% from baseline. Smooth degradation, no cliff."
} | {
"date": "2026-05-14T00:00:00",
"run_type": "warm_server"
} |
google/gemma-4-E2B-it | gemma-4-e2b | google | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 65536,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 72.1,
"ttft_seconds": null,
"tokens_generated": 512,
"vram_used_gb": 2.61,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-48k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~48K context fill (37530 prompt tokens). Prompt eval 4926 tok/s. VRAM 2670 MiB. -39% from baseline. Still 5518 MiB free. Smooth scaling."
} | {
"date": "2026-05-14T00:00:00",
"run_type": "warm_server"
} |
google/gemma-4-E2B-it | gemma-4-e2b | google | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 65536,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 116.8,
"ttft_seconds": null,
"tokens_generated": 2048,
"vram_used_gb": 2.6,
"ram_used_gb_peak": null,
"prompt_id": "sustained-throughput-2048"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "Sustained throughput test: 2048 tokens generated at low context depth (61 prompt tokens). Consistent speed across long output. Quality: 4/6 pass, 2/6 partial (logic puzzle confusion, hedged hallucination). PLE incomplete in llama.cpp."
} | {
"date": "2026-05-14T00:00:00",
"run_type": "warm_server"
} |
liquid-ai/LFM2-24B-A2B | lfm2-24b-a2b | liquid-ai | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 16384,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 37.28,
"ttft_seconds": null,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=30 (10 layers experts on GPU, 30 on CPU). ngl=999, np=1, threads=6. VRAM projected 4520 MiB / 7063 MiB free (2540 MiB headroom). 64 experts/layer, 4 active. Comfortable but underutilising GPU — lots of headroom remaining."
} | {
"date": "2026-05-13T00:00:00",
"run_type": "cold_server"
} |
liquid-ai/LFM2-24B-A2B | lfm2-24b-a2b | liquid-ai | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 16384,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 44.05,
"ttft_seconds": null,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=26 (14 layers experts on GPU, 26 on CPU). ngl=999, np=1, threads=6. VRAM projected 5866 MiB / 7063 MiB free (1194 MiB headroom). Good balance of speed and headroom."
} | {
"date": "2026-05-13T00:00:00",
"run_type": "cold_server"
} |
liquid-ai/LFM2-24B-A2B | lfm2-24b-a2b | liquid-ai | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 16384,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 47.52,
"ttft_seconds": null,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=24 (16 layers experts on GPU, 24 on CPU). ngl=999, np=1, threads=6. VRAM projected 6563 MiB / 7063 MiB free (498 MiB headroom). Safe practical config with good speed."
} | {
"date": "2026-05-13T00:00:00",
"run_type": "cold_server"
} |
liquid-ai/LFM2-24B-A2B | lfm2-24b-a2b | liquid-ai | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 16384,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 49.79,
"ttft_seconds": null,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=22 SWEET SPOT at 16K (18 layers experts on GPU, 22 on CPU). ngl=999, np=1, threads=6. VRAM projected 7261 MiB / 7063 MiB free (-200 MiB over budget). Over-commits but no cliff — LFM2 handles VRAM pressure gracefully. Hybrid SSM+conv architecture: 1... | {
"date": "2026-05-13T00:00:00",
"run_type": "cold_server"
} |
liquid-ai/LFM2-24B-A2B | lfm2-24b-a2b | liquid-ai | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 52.23,
"ttft_seconds": null,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=22 BEST OVERALL CONFIG at 32K (18 layers experts on GPU, 22 on CPU). ngl=999, np=1, threads=6. VRAM projected 7479 MiB / 7063 MiB free (-418 MiB over budget). 32K context essentially free vs 16K (52.23 vs 49.79 — same counterintuitive pattern as Qw... | {
"date": "2026-05-13T00:00:00",
"run_type": "cold_server"
} |
liquid-ai/LFM2-24B-A2B | lfm2-24b-a2b | liquid-ai | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 65536,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 37.56,
"ttft_seconds": null,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=22 at 65K (18 layers experts on GPU, 22 on CPU). ngl=999, np=1, threads=6. VRAM projected 7915 MiB / 7063 MiB free (-854 MiB over budget). Gentle degradation: -28% vs 32K sweet spot (37.56 vs 52.23). Compares favourably to Qwen at 65K (-51%) and Ge... | {
"date": "2026-05-13T00:00:00",
"run_type": "cold_server"
} |
mistralai/Mistral-7B-Instruct-v0.3 | mistral-7b-v0.3 | mistralai | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 56.4,
"ttft_seconds": null,
"tokens_generated": 512,
"vram_used_gb": 6.56,
"ram_used_gb_peak": null,
"prompt_id": "tcp-udp-explanation"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "Baseline at low context depth (37 prompt tokens). ngl=999, np=1, threads=6. Full GPU offload, 6722 MiB VRAM used, 1466 MiB free. Dense 7.3B all-active. VRAM-hungry but fast. No thinking mode."
} | {
"date": "2026-05-14T00:00:00",
"run_type": "cold_server"
} |
mistralai/Mistral-7B-Instruct-v0.3 | mistral-7b-v0.3 | mistralai | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 43.4,
"ttft_seconds": null,
"tokens_generated": 486,
"vram_used_gb": 6.59,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-8k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~8K context fill (7273 prompt tokens). Prompt eval 3070 tok/s. VRAM 6752 MiB. -23% from baseline."
} | {
"date": "2026-05-14T00:00:00",
"run_type": "warm_server"
} |
mistralai/Mistral-7B-Instruct-v0.3 | mistral-7b-v0.3 | mistralai | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 35.7,
"ttft_seconds": null,
"tokens_generated": 414,
"vram_used_gb": 6.62,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-16k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~16K context fill (14523 prompt tokens). Prompt eval 2610 tok/s. VRAM 6780 MiB. -37% from baseline."
} | {
"date": "2026-05-14T00:00:00",
"run_type": "warm_server"
} |
mistralai/Mistral-7B-Instruct-v0.3 | mistral-7b-v0.3 | mistralai | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 30.5,
"ttft_seconds": null,
"tokens_generated": 103,
"vram_used_gb": 6.65,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-24k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~24K context fill (21777 prompt tokens). Prompt eval 2145 tok/s. VRAM 6810 MiB. -46% from baseline."
} | {
"date": "2026-05-14T00:00:00",
"run_type": "warm_server"
} |
mistralai/Mistral-7B-Instruct-v0.3 | mistral-7b-v0.3 | mistralai | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 26.4,
"ttft_seconds": null,
"tokens_generated": 116,
"vram_used_gb": 6.68,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-32k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~32K context fill (29027 prompt tokens). Prompt eval 1740 tok/s. VRAM 6835 MiB. -53% from baseline. Steep degradation, all 32 layers full attention."
} | {
"date": "2026-05-14T00:00:00",
"run_type": "warm_server"
} |
mistralai/Mistral-7B-Instruct-v0.3 | mistral-7b-v0.3 | mistralai | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 56.2,
"ttft_seconds": null,
"tokens_generated": 798,
"vram_used_gb": 6.62,
"ram_used_gb_peak": null,
"prompt_id": "sustained-throughput-2048"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "Sustained throughput test: asked for 2048, got 798 (hit EOS). Speed held at 56.2 tok/s. No thinking mode. Quality: 5/6 pass, 1/6 partial. Best quality score. Confident hallucinations: claims fictional researchers are 'notable'."
} | {
"date": "2026-05-14T00:00:00",
"run_type": "warm_server"
} |
zai-org/GLM-4.7-Flash | glm-4.7-flash | zai-org | Q4_K_M | GGUF | moe_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": false,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": 35,
"ngl"... | {
"tok_per_sec": 29.9,
"ttft_seconds": null,
"tokens_generated": 512,
"vram_used_gb": 6.97,
"ram_used_gb_peak": null,
"prompt_id": "tcp-udp-explanation"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "Baseline at low context depth (34 prompt tokens). ngl=999, ncmoe=35, np=1, threads=6. MoE offload, 7140 MiB VRAM used. deepseek2 arch (MoE + MLA). 30B total, ~3B active per token. 64 experts + 1 shared, 4 active. Flash attention OFF (required for MLA). -... | {
"date": "2026-05-15T00:00:00",
"run_type": "cold_server"
} |
zai-org/GLM-4.7-Flash | glm-4.7-flash | zai-org | Q4_K_M | GGUF | moe_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": false,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": 35,
"ngl"... | {
"tok_per_sec": 29,
"ttft_seconds": null,
"tokens_generated": 256,
"vram_used_gb": 6.98,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-8k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~8K context fill (6020 prompt tokens). Prompt eval 450.5 tok/s. VRAM 7148 MiB. -3.0% from baseline. MLA keeps KV cache small."
} | {
"date": "2026-05-15T00:00:00",
"run_type": "warm_server"
} |
zai-org/GLM-4.7-Flash | glm-4.7-flash | zai-org | Q4_K_M | GGUF | moe_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": false,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": 35,
"ngl"... | {
"tok_per_sec": 27.1,
"ttft_seconds": null,
"tokens_generated": 256,
"vram_used_gb": 6.99,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-16k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~16K context fill (12020 prompt tokens). Prompt eval 437.1 tok/s. VRAM 7154 MiB. -9.4% from baseline."
} | {
"date": "2026-05-15T00:00:00",
"run_type": "warm_server"
} |
zai-org/GLM-4.7-Flash | glm-4.7-flash | zai-org | Q4_K_M | GGUF | moe_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": false,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": 35,
"ngl"... | {
"tok_per_sec": 25.5,
"ttft_seconds": null,
"tokens_generated": 256,
"vram_used_gb": 6.99,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-24k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~24K context fill (18020 prompt tokens). Prompt eval 422.3 tok/s. VRAM 7160 MiB. -14.7% from baseline. Flattest context degradation of all models tested. MLA compresses KV cache."
} | {
"date": "2026-05-15T00:00:00",
"run_type": "warm_server"
} |
zai-org/GLM-4.7-Flash | glm-4.7-flash | zai-org | Q4_K_M | GGUF | moe_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": false,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": 35,
"ngl"... | {
"tok_per_sec": 29.5,
"ttft_seconds": null,
"tokens_generated": 2048,
"vram_used_gb": 6.99,
"ram_used_gb_peak": null,
"prompt_id": "sustained-throughput-2048"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "Sustained throughput test: asked for 2048, got 2048 (full budget). Speed held at 29.5 tok/s. Thinking mode active. Quality: 4/6 pass, 2/6 fail. Code gen fail (reasoning consumed all tokens). Hallucination fail (fabricated complete paper with specific det... | {
"date": "2026-05-15T00:00:00",
"run_type": "warm_server"
} |
zai-org/GLM-4.7-Flash | glm-4.7-flash | zai-org | Q4_K_M | GGUF | moe_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": false,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": 35,
"ngl"... | {
"tok_per_sec": 29.9,
"ttft_seconds": null,
"tokens_generated": 786,
"vram_used_gb": 6.99,
"ram_used_gb_peak": null,
"prompt_id": "ncmoe-sweep"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe sweep results: ncmoe40=25.8tok/s(4951MiB), ncmoe35=28.8tok/s(6704MiB, sweet spot), ncmoe32=32.1tok/s(7789MiB, tight), ncmoe30=24.6tok/s(7823MiB, VRAM cliff). Flash attention OFF required for MLA arch."
} | {
"date": "2026-05-15T00:00:00",
"run_type": "warm_server"
} |
Qwen/Qwen3-8B | qwen3-8b | Qwen | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 50.8,
"ttft_seconds": null,
"tokens_generated": 512,
"vram_used_gb": 7.65,
"ram_used_gb_peak": null,
"prompt_id": "tcp-udp-explanation"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "Baseline at low context depth (37 prompt tokens). ngl=999, np=1, threads=6. Full GPU offload, 7836 MiB VRAM used. Dense 8.2B all-active. VRAM tight at 32K: model 4.7 GB + KV cache 2.4 GB. Compute-bound not memory-bound."
} | {
"date": "2026-05-16T00:00:00",
"run_type": "cold_server"
} |
Qwen/Qwen3-8B | qwen3-8b | Qwen | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 44.8,
"ttft_seconds": null,
"tokens_generated": 256,
"vram_used_gb": 7.68,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-8k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~8K context fill (6023 prompt tokens). Prompt eval 2943.9 tok/s. VRAM 7862 MiB. -11.8% from baseline."
} | {
"date": "2026-05-16T00:00:00",
"run_type": "warm_server"
} |
Qwen/Qwen3-8B | qwen3-8b | Qwen | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 40.4,
"ttft_seconds": null,
"tokens_generated": 256,
"vram_used_gb": 7.7,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-16k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~16K context fill (12023 prompt tokens). Prompt eval 2314.1 tok/s. VRAM 7884 MiB. -20.5% from baseline."
} | {
"date": "2026-05-16T00:00:00",
"run_type": "warm_server"
} |
Qwen/Qwen3-8B | qwen3-8b | Qwen | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 36.6,
"ttft_seconds": null,
"tokens_generated": 256,
"vram_used_gb": 7.72,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-24k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~24K context fill (18023 prompt tokens). Prompt eval 1713.3 tok/s. VRAM 7906 MiB. -27.9% from baseline."
} | {
"date": "2026-05-16T00:00:00",
"run_type": "warm_server"
} |
Qwen/Qwen3-8B | qwen3-8b | Qwen | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 50,
"ttft_seconds": null,
"tokens_generated": 2048,
"vram_used_gb": 7.72,
"ram_used_gb_peak": null,
"prompt_id": "sustained-throughput-2048"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "Sustained throughput test: 2048 tokens @ 50.0 tok/s. Thinking mode active. Quality: 4/6 pass, 1/6 partial, 1/6 fail. Code gen fail (reasoning consumed all tokens). Hallucination pass (first model to correctly identify fictional study)."
} | {
"date": "2026-05-16T00:00:00",
"run_type": "warm_server"
} |
meta-llama/Llama-3.2-1B-Instruct | llama-3.2-1b | meta-llama | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 228.2,
"ttft_seconds": null,
"tokens_generated": 512,
"vram_used_gb": 2.07,
"ram_used_gb_peak": null,
"prompt_id": "tcp-udp-explanation"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "Baseline at low context depth (64 prompt tokens). ngl=999, np=1, threads=6. Full GPU offload, 2123 MiB VRAM used, 5829 MiB free. Dense 1.24B all-active. Speed ceiling for this hardware. No thinking mode."
} | {
"date": "2026-05-16T00:00:00",
"run_type": "cold_server"
} |
meta-llama/Llama-3.2-1B-Instruct | llama-3.2-1b | meta-llama | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 210.7,
"ttft_seconds": null,
"tokens_generated": 256,
"vram_used_gb": 2.08,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-8k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~8K context fill (6050 prompt tokens). Prompt eval 16791 tok/s. VRAM 2135 MiB. -7.7% from baseline."
} | {
"date": "2026-05-16T00:00:00",
"run_type": "warm_server"
} |
meta-llama/Llama-3.2-1B-Instruct | llama-3.2-1b | meta-llama | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 189.8,
"ttft_seconds": null,
"tokens_generated": 251,
"vram_used_gb": 2.1,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-16k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~16K context fill (12050 prompt tokens). Prompt eval 12570.8 tok/s. VRAM 2147 MiB. -16.8% from baseline."
} | {
"date": "2026-05-16T00:00:00",
"run_type": "warm_server"
} |
meta-llama/Llama-3.2-1B-Instruct | llama-3.2-1b | meta-llama | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 171.4,
"ttft_seconds": null,
"tokens_generated": 256,
"vram_used_gb": 2.11,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-24k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~24K context fill (18050 prompt tokens). Prompt eval 9950.4 tok/s. VRAM 2159 MiB. -24.9% from baseline."
} | {
"date": "2026-05-16T00:00:00",
"run_type": "warm_server"
} |
meta-llama/Llama-3.2-1B-Instruct | llama-3.2-1b | meta-llama | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 235.5,
"ttft_seconds": null,
"tokens_generated": 1663,
"vram_used_gb": 2.11,
"ram_used_gb_peak": null,
"prompt_id": "sustained-throughput-2048"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "Sustained throughput: asked 2048, got 1663 (hit EOS). Speed held at 235.5 tok/s. No thinking mode. Quality: 4/6 pass, 2/6 fail. JSON fences fail, logic puzzle fail (missed riddle entirely)."
} | {
"date": "2026-05-16T00:00:00",
"run_type": "warm_server"
} |
Qwen/Qwen3.6-27B | qwen3.6-27b-dense | Qwen | Q4_K_M | GGUF | dense_partial_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "partial",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
... | {
"tok_per_sec": 3.28,
"ttft_seconds": null,
"tokens_generated": 512,
"vram_used_gb": 6.63,
"ram_used_gb_peak": 12.1,
"prompt_id": "tcp-udp-explanation"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "Baseline at low context depth (19 prompt tokens). ngl=20 of 64 layers, -c 32768. 27B all-active dense hybrid SSM+attention. Partial GPU offload: 5547 MiB model on GPU, 11163 MiB on host. VRAM total 6784 MiB. Prompt eval 7.9 tok/s. Brutally slow decode: C... | {
"date": "2026-05-18T00:00:00",
"run_type": "cold_server"
} |
Qwen/Qwen3.6-27B | qwen3.6-27b-dense | Qwen | Q4_K_M | GGUF | dense_partial_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "partial",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
... | {
"tok_per_sec": 2.96,
"ttft_seconds": null,
"tokens_generated": 256,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-8k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~8K context fill (6363 prompt tokens). Prompt eval 272.8 tok/s. -9.8% from baseline. DeltaNet layers (48/64) don't scale with context, degradation from 16 attention layers only."
} | {
"date": "2026-05-18T00:00:00",
"run_type": "warm_server"
} |
Qwen/Qwen3.6-27B | qwen3.6-27b-dense | Qwen | Q4_K_M | GGUF | dense_partial_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "partial",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
... | {
"tok_per_sec": 2.62,
"ttft_seconds": null,
"tokens_generated": 256,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-16k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~16K context fill (12703 prompt tokens). Prompt eval 284.5 tok/s. -20.1% from baseline."
} | {
"date": "2026-05-18T00:00:00",
"run_type": "warm_server"
} |
Qwen/Qwen3.6-27B | qwen3.6-27b-dense | Qwen | Q4_K_M | GGUF | dense_partial_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "partial",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
... | {
"tok_per_sec": 2.12,
"ttft_seconds": null,
"tokens_generated": 256,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-24k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~24K context fill (19045 prompt tokens). Prompt eval 280.1 tok/s. -35.4% from baseline. Worst context degradation in benchmark set."
} | {
"date": "2026-05-18T00:00:00",
"run_type": "warm_server"
} |
Qwen/Qwen3.6-27B | qwen3.6-27b-dense | Qwen | Q4_K_M | GGUF | dense_partial_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "partial",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
... | {
"tok_per_sec": 3.15,
"ttft_seconds": null,
"tokens_generated": 2048,
"vram_used_gb": 6.63,
"ram_used_gb_peak": null,
"prompt_id": "sustained-throughput-2048"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "Sustained throughput test: 2048 max tokens. Average decode across 6 quality tests: 3.14 tok/s. Quality: 5/6 pass, 0 partial, 1 fail. Best quality of all 9 models. Best hallucination resistance: firm refusal of fictional study. Code gen fail: thinking con... | {
"date": "2026-05-18T00:00:00",
"run_type": "warm_server"
} |
openai/gpt-oss-20b | gpt-oss-20b | openai | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 22,
"ttft_seconds": null,
"tokens_generated": 1024,
"vram_used_gb": 2.14,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=30 (all 24 MoE layers' experts on CPU). 21B total, 3.6B active, 32 experts/layer, 4 active. OpenAI Apache 2.0. Native MXFP4 FFN weights. Maximum headroom config (6 GB free). This is the config used for agentic coding (10/10 tasks completed via Pi a... | {
"date": "2026-05-20T00:00:00",
"run_type": "cold_server"
} |
openai/gpt-oss-20b | gpt-oss-20b | openai | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 25.7,
"ttft_seconds": null,
"tokens_generated": 1024,
"vram_used_gb": 3.78,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=20 (4 MoE layers' experts on GPU, 20 on CPU). ~407 MiB per MoE layer on GPU. tok/s estimated from client-side measurement (24.86) with ~3.4% HTTP overhead adjustment."
} | {
"date": "2026-05-20T00:00:00",
"run_type": "cold_server"
} |
openai/gpt-oss-20b | gpt-oss-20b | openai | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 39.45,
"ttft_seconds": null,
"tokens_generated": 1024,
"vram_used_gb": 7.02,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=12 SWEET SPOT at 32K (12 MoE layers' experts on GPU, 12 on CPU). 7024 MiB VRAM with 1164 MiB headroom. Best balance of speed and safety. Architecture: 24 layers, 32 experts/layer, 4 active, n_embd=2880. Per-layer expert cost ~407 MiB on GPU."
} | {
"date": "2026-05-20T00:00:00",
"run_type": "cold_server"
} |
openai/gpt-oss-20b | gpt-oss-20b | openai | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 16384,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 40.6,
"ttft_seconds": null,
"tokens_generated": 1024,
"vram_used_gb": 6.91,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=12 at 16K context. Context scaling nearly flat vs 32K (40.60 vs 39.45 tok/s). KV cache is small relative to expert weights at this model size."
} | {
"date": "2026-05-20T00:00:00",
"run_type": "cold_server"
} |
openai/gpt-oss-20b | gpt-oss-20b | openai | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 46.64,
"ttft_seconds": null,
"tokens_generated": 1024,
"vram_used_gb": 7.83,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=10 MAX SPEED (14 MoE layers' experts on GPU, 10 on CPU). 7832 MiB VRAM with only 356 MiB headroom. Fastest config but tight on VRAM. No CUDA cliff observed despite 7.8 GB usage."
} | {
"date": "2026-05-20T00:00:00",
"run_type": "cold_server"
} |
google/gemma-4-E4B-it | gemma-4-e4b | google | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 64.12,
"ttft_seconds": null,
"tokens_generated": 1024,
"vram_used_gb": 4.13,
"ram_used_gb_peak": null,
"prompt_id": "tcp-udp-explanation"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "Baseline at low context depth (44 prompt tokens). ngl=999, np=1, threads=6. Full GPU offload, 4225 MiB VRAM used, 3963 MiB free. Dense model with PLE, no -ncmoe needed. Built-in thinking mode. 6/6 quality pass."
} | {
"date": "2026-05-21T00:00:00",
"run_type": "cold_server"
} |
google/gemma-4-E4B-it | gemma-4-e4b | google | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 58.1,
"ttft_seconds": null,
"tokens_generated": 512,
"vram_used_gb": 4.13,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-8k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~8K context fill (8040 prompt tokens). Prompt eval 4312 tok/s. VRAM 4226 MiB. -9% from baseline."
} | {
"date": "2026-05-21T00:00:00",
"run_type": "warm_server"
} |
google/gemma-4-E4B-it | gemma-4-e4b | google | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 50.35,
"ttft_seconds": null,
"tokens_generated": 512,
"vram_used_gb": 4.13,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-16k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~16K context fill (16040 prompt tokens). Prompt eval 3839 tok/s. VRAM 4226 MiB. -21% from baseline."
} | {
"date": "2026-05-21T00:00:00",
"run_type": "warm_server"
} |
google/gemma-4-E4B-it | gemma-4-e4b | google | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 39.77,
"ttft_seconds": null,
"tokens_generated": 512,
"vram_used_gb": 4.22,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-32k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~32K context fill (32040 prompt tokens). Prompt eval 2989 tok/s. VRAM 4315 MiB. -38% from baseline. Smooth degradation, no cliff."
} | {
"date": "2026-05-21T00:00:00",
"run_type": "warm_server"
} |
google/gemma-4-E4B-it | gemma-4-e4b | google | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 63.91,
"ttft_seconds": null,
"tokens_generated": 2048,
"vram_used_gb": 4.19,
"ram_used_gb_peak": null,
"prompt_id": "sustained-throughput-2048"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "Sustained throughput test: 2048 tokens generated at low context depth (55 prompt tokens). Consistent speed across long output. Quality: 6/6 pass (best on leaderboard). PLE may not be fully implemented in llama.cpp."
} | {
"date": "2026-05-21T00:00:00",
"run_type": "warm_server"
} |
Qwen/Qwen3.6-35B-A3B | qwen3.6-35b-a3b | qwen | Q4_K_S | GGUF | moe_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "Windows 11"
} | {
"engine": "LM Studio",
"context_length": 16384,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": true,
"temperature": 0.7,
"top_p": 0.9,
"ncmoe": null,
"ngl": null
} | {
"tok_per_sec": 7.4,
"ttft_seconds": 1.75,
"tokens_generated": 761,
"vram_used_gb": null,
"ram_used_gb_peak": 31,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": 4,
"hit_length_target": true,
"notes": "RAM-ceilinged on 32GB system. Reproduces 5.5x slower than the public X-post claim of 41 tok/sec — same author's thread requires 64GB RAM as the unstated prerequisite. GGUF repackaged by bartowski."
} | {
"date": "2026-05-06T00:00:00",
"run_type": "warm"
} |
Qwen/Qwen3.6-35B-A3B | qwen3.6-35b-a3b | qwen | Q4_K_S | GGUF | moe_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "Windows 11"
} | {
"engine": "LM Studio",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": true,
"temperature": 0.7,
"top_p": 0.9,
"ncmoe": null,
"ngl": null
} | {
"tok_per_sec": 7.43,
"ttft_seconds": 1.85,
"tokens_generated": 796,
"vram_used_gb": null,
"ram_used_gb_peak": 31,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": 4,
"hit_length_target": true,
"notes": "Doubling context from 16K to 32K barely moved tok/sec (7.40 → 7.43). Flat curve = constant overhead dominates, not KV-attention-scan cost. Consistent with disk-paging bottleneck rather than RAM-bandwidth bottleneck. GGUF repackaged by bartowski."
} | {
"date": "2026-05-06T00:00:00",
"run_type": "warm"
} |
Qwen/Qwen3.6-35B-A3B | qwen3.6-35b-a3b | qwen | IQ4_XS | GGUF | moe_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "Windows 11"
} | {
"engine": "LM Studio",
"context_length": 16384,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": true,
"temperature": 0.7,
"top_p": 0.9,
"ncmoe": null,
"ngl": null
} | {
"tok_per_sec": 7.4,
"ttft_seconds": null,
"tokens_generated": null,
"vram_used_gb": null,
"ram_used_gb_peak": 31,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "Speed-only run, no quality eval captured. Smaller quant (~2GB less than Q4_K_S, 18.81GB vs 20.59GB) tested to isolate quant size as variable. Same tok/sec confirms quant size is not the lever — RAM ceiling is the binding constraint regardless of weight f... | {
"date": "2026-05-06T00:00:00",
"run_type": "warm"
} |
Qwen/Qwen3.6-35B-A3B | qwen3.6-35b-a3b | qwen | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (llama.cpp b9049-2496f9c14)",
"context_length": 16384,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 29.7,
"ttft_seconds": 2.44,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=32 (8 layers experts on GPU, 32 on CPU). ngl=999, threads=6. Prompt: 20.46 tok/s. Total: 36.9s. VRAM ~6649 MiB / 8187 MiB. CPU-mapped experts: 16581 MiB. Quantized by Unsloth (UD). Baseline partial offload — clean, no VRAM pressure. 4x faster than ... | {
"date": "2026-05-07T00:00:00",
"run_type": "cold_server"
} |
Qwen/Qwen3.6-35B-A3B | qwen3.6-35b-a3b | qwen | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (llama.cpp b9049-2496f9c14)",
"context_length": 16384,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 32,
"ttft_seconds": 1.18,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=30 (10 layers experts on GPU, 30 on CPU). ngl=999, threads=6. Prompt: 42.48 tok/s. Total: 33.2s. Sweet spot for 16K context. 7.7% faster decode than ncmoe 32. 4.3x faster than full offload."
} | {
"date": "2026-05-07T00:00:00",
"run_type": "cold_server"
} |
Qwen/Qwen3.6-35B-A3B | qwen3.6-35b-a3b | qwen | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (llama.cpp b9049-2496f9c14)",
"context_length": 16384,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 16.33,
"ttft_seconds": 2.05,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=28 (12 layers experts on GPU, 28 on CPU). ngl=999, threads=6. Prompt: 24.38 tok/s. Total: 64.7s. VRAM pressure triggered — CUDA page faults through PCIe halved throughput. Sharp cliff, not gradual."
} | {
"date": "2026-05-07T00:00:00",
"run_type": "cold_server"
} |
Qwen/Qwen3.6-35B-A3B | qwen3.6-35b-a3b | qwen | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (llama.cpp b9049-2496f9c14)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 35.36,
"ttft_seconds": 1.44,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=30 (10 layers experts on GPU, 30 on CPU). ngl=999, threads=6. Prompt: 34.72 tok/s. Total: 30.4s. Best overall config. Hybrid SSM+attention means KV cache only scales for 10/40 attention layers — 32K is nearly free. 4.8x faster than full offload. Re... | {
"date": "2026-05-07T00:00:00",
"run_type": "cold_server"
} |
Qwen/Qwen3.6-35B-A3B | qwen3.6-35b-a3b | qwen | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (llama.cpp b9049-2496f9c14)",
"context_length": 65536,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 17.41,
"ttft_seconds": 2.83,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=30 (10 layers experts on GPU, 30 on CPU). ngl=999, threads=6. Prompt: 17.70 tok/s. Total: 61.7s. VRAM cliff at 65K — KV cache for 10 attention layers at q8_0 grows to ~680 MiB, pushing total GPU past ~7 GB. Same ~50% throughput drop as ncmoe 28 at ... | {
"date": "2026-05-07T00:00:00",
"run_type": "cold_server"
} |
google/gemma-4-26B-A4B | gemma-4-26b-a4b | google | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 16384,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 27.92,
"ttft_seconds": 1.45,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=24 (6 layers experts on GPU, 24 on CPU). ngl=999, np=1, threads=6. Prompt: 43.54 tok/s. Total: 38.1s. VRAM projected 6119 MiB / 7063 MiB free (944 MiB headroom). 128 experts/layer, 8+1 shared active. Comfortable headroom but fewer expert layers on ... | {
"date": "2026-05-13T00:00:00",
"run_type": "cold_server"
} |
google/gemma-4-26B-A4B | gemma-4-26b-a4b | google | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 16384,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 29.19,
"ttft_seconds": 1.59,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=23 (7 layers experts on GPU, 23 on CPU). ngl=999, np=1, threads=6. Prompt: 39.67 tok/s. Total: 36.7s. VRAM projected 6573 MiB / 7063 MiB free (490 MiB headroom). SWEET SPOT at 16K. Best decode speed with safe VRAM margin."
} | {
"date": "2026-05-13T00:00:00",
"run_type": "cold_server"
} |
google/gemma-4-26B-A4B | gemma-4-26b-a4b | google | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 16384,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 28.34,
"ttft_seconds": 1.57,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=22 (8 layers experts on GPU, 22 on CPU). ngl=999, np=1, threads=6. Prompt: 40.08 tok/s. Total: 37.7s. VRAM projected 7026 MiB / 7063 MiB free (37 MiB headroom). Razor-thin margin — works but risky for sustained use."
} | {
"date": "2026-05-13T00:00:00",
"run_type": "cold_server"
} |
google/gemma-4-26B-A4B | gemma-4-26b-a4b | google | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 16384,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 24.5,
"ttft_seconds": 1.82,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=20 (10 layers experts on GPU, 20 on CPU). ngl=999, np=1, threads=6. Prompt: 34.54 tok/s. Total: 43.6s. VRAM projected 7934 MiB / 7063 MiB free (-871 MiB over budget). CUDA page faults, 16% speed drop. Gentler cliff than Qwen (-50% at same over-budg... | {
"date": "2026-05-13T00:00:00",
"run_type": "cold_server"
} |
google/gemma-4-26B-A4B | gemma-4-26b-a4b | google | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 29.27,
"ttft_seconds": 1.14,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=23 (7 layers experts on GPU, 23 on CPU). ngl=999, np=1, threads=6. Prompt: 55.31 tok/s. Total: 36.1s. VRAM projected 6743 MiB / 7063 MiB free (320 MiB headroom). BEST OVERALL CONFIG. 32K context nearly free thanks to 5:1 SWA-to-full attention patte... | {
"date": "2026-05-13T00:00:00",
"run_type": "cold_server"
} |
google/gemma-4-26B-A4B | gemma-4-26b-a4b | google | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 65536,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 25.78,
"ttft_seconds": 1.16,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=23 (7 layers experts on GPU, 23 on CPU). ngl=999, np=1, threads=6. Prompt: 54.12 tok/s. Total: 40.9s. VRAM projected 7083 MiB / 7063 MiB free (-20 MiB over budget). Gentle degradation (-12% vs 16K). SWA pattern keeps KV at 839 MiB. No catastrophic ... | {
"date": "2026-05-13T00:00:00",
"run_type": "cold_server"
} |
google/gemma-4-E4B-it | gemma-4-e4b | google | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 64.12,
"ttft_seconds": null,
"tokens_generated": 1024,
"vram_used_gb": 4.13,
"ram_used_gb_peak": null,
"prompt_id": "tcp-udp-explanation"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "Baseline at low context depth (44 prompt tokens). ngl=999, np=1, threads=6. Full GPU offload, 4225 MiB VRAM used, 3963 MiB free. Dense model with PLE, no -ncmoe needed. Built-in thinking mode. 6/6 quality pass."
} | {
"date": "2026-05-21T00:00:00",
"run_type": "cold_server"
} |
google/gemma-4-E4B-it | gemma-4-e4b | google | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 58.1,
"ttft_seconds": null,
"tokens_generated": 512,
"vram_used_gb": 4.13,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-8k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~8K context fill (8040 prompt tokens). Prompt eval 4312 tok/s. VRAM 4226 MiB. -9% from baseline."
} | {
"date": "2026-05-21T00:00:00",
"run_type": "warm_server"
} |
google/gemma-4-E4B-it | gemma-4-e4b | google | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 50.35,
"ttft_seconds": null,
"tokens_generated": 512,
"vram_used_gb": 4.13,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-16k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~16K context fill (16040 prompt tokens). Prompt eval 3839 tok/s. VRAM 4226 MiB. -21% from baseline."
} | {
"date": "2026-05-21T00:00:00",
"run_type": "warm_server"
} |
google/gemma-4-E4B-it | gemma-4-e4b | google | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 39.77,
"ttft_seconds": null,
"tokens_generated": 512,
"vram_used_gb": 4.22,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-32k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~32K context fill (32040 prompt tokens). Prompt eval 2989 tok/s. VRAM 4315 MiB. -38% from baseline. Smooth degradation, no cliff."
} | {
"date": "2026-05-21T00:00:00",
"run_type": "warm_server"
} |
google/gemma-4-E4B-it | gemma-4-e4b | google | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 63.91,
"ttft_seconds": null,
"tokens_generated": 2048,
"vram_used_gb": 4.19,
"ram_used_gb_peak": null,
"prompt_id": "sustained-throughput-2048"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "Sustained throughput test: 2048 tokens generated at low context depth (55 prompt tokens). Consistent speed across long output. Quality: 6/6 pass (best on leaderboard). PLE may not be fully implemented in llama.cpp."
} | {
"date": "2026-05-21T00:00:00",
"run_type": "warm_server"
} |
google/gemma-4-E2B-it | gemma-4-e2b | google | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 65536,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 117.8,
"ttft_seconds": null,
"tokens_generated": 512,
"vram_used_gb": 2.53,
"ram_used_gb_peak": null,
"prompt_id": "tcp-udp-explanation"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "Baseline at low context depth (44 prompt tokens). ngl=999, np=1, threads=6. Full GPU offload, 2593 MiB VRAM used, 5595 MiB free. Dense model, no -ncmoe needed. Built-in thinking mode (reasoning_content). PLE not fully implemented in llama.cpp."
} | {
"date": "2026-05-14T00:00:00",
"run_type": "cold_server"
} |
google/gemma-4-E2B-it | gemma-4-e2b | google | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 65536,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 113.1,
"ttft_seconds": null,
"tokens_generated": 512,
"vram_used_gb": 2.54,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-8k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~8K context fill (6280 prompt tokens). Prompt eval 7761 tok/s. VRAM 2605 MiB. Minimal degradation from baseline."
} | {
"date": "2026-05-14T00:00:00",
"run_type": "warm_server"
} |
google/gemma-4-E2B-it | gemma-4-e2b | google | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 65536,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 107.3,
"ttft_seconds": null,
"tokens_generated": 512,
"vram_used_gb": 2.55,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-16k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~16K context fill (12530 prompt tokens). Prompt eval 7317 tok/s. VRAM 2617 MiB. -9% from baseline."
} | {
"date": "2026-05-14T00:00:00",
"run_type": "warm_server"
} |
google/gemma-4-E2B-it | gemma-4-e2b | google | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 65536,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 87.7,
"ttft_seconds": null,
"tokens_generated": 512,
"vram_used_gb": 2.58,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-32k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~32K context fill (25030 prompt tokens). Prompt eval 6111 tok/s. VRAM 2641 MiB. -26% from baseline. Smooth degradation, no cliff."
} | {
"date": "2026-05-14T00:00:00",
"run_type": "warm_server"
} |
google/gemma-4-E2B-it | gemma-4-e2b | google | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 65536,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 72.1,
"ttft_seconds": null,
"tokens_generated": 512,
"vram_used_gb": 2.61,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-48k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~48K context fill (37530 prompt tokens). Prompt eval 4926 tok/s. VRAM 2670 MiB. -39% from baseline. Still 5518 MiB free. Smooth scaling."
} | {
"date": "2026-05-14T00:00:00",
"run_type": "warm_server"
} |
google/gemma-4-E2B-it | gemma-4-e2b | google | Q4_K_M | GGUF | full_gpu_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 65536,
"gpu_offload": "max",
"expert_offload_to_cpu": false,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": null,
"ng... | {
"tok_per_sec": 116.8,
"ttft_seconds": null,
"tokens_generated": 2048,
"vram_used_gb": 2.6,
"ram_used_gb_peak": null,
"prompt_id": "sustained-throughput-2048"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "Sustained throughput test: 2048 tokens generated at low context depth (61 prompt tokens). Consistent speed across long output. Quality: 4/6 pass, 2/6 partial (logic puzzle confusion, hedged hallucination). PLE incomplete in llama.cpp."
} | {
"date": "2026-05-14T00:00:00",
"run_type": "warm_server"
} |
zai-org/GLM-4.7-Flash | glm-4.7-flash | zai-org | Q4_K_M | GGUF | moe_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": false,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": 35,
"ngl"... | {
"tok_per_sec": 29.9,
"ttft_seconds": null,
"tokens_generated": 512,
"vram_used_gb": 6.97,
"ram_used_gb_peak": null,
"prompt_id": "tcp-udp-explanation"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "Baseline at low context depth (34 prompt tokens). ngl=999, ncmoe=35, np=1, threads=6. MoE offload, 7140 MiB VRAM used. deepseek2 arch (MoE + MLA). 30B total, ~3B active per token. 64 experts + 1 shared, 4 active. Flash attention OFF (required for MLA). -... | {
"date": "2026-05-15T00:00:00",
"run_type": "cold_server"
} |
zai-org/GLM-4.7-Flash | glm-4.7-flash | zai-org | Q4_K_M | GGUF | moe_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": false,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": 35,
"ngl"... | {
"tok_per_sec": 29,
"ttft_seconds": null,
"tokens_generated": 256,
"vram_used_gb": 6.98,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-8k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~8K context fill (6020 prompt tokens). Prompt eval 450.5 tok/s. VRAM 7148 MiB. -3.0% from baseline. MLA keeps KV cache small."
} | {
"date": "2026-05-15T00:00:00",
"run_type": "warm_server"
} |
zai-org/GLM-4.7-Flash | glm-4.7-flash | zai-org | Q4_K_M | GGUF | moe_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": false,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": 35,
"ngl"... | {
"tok_per_sec": 27.1,
"ttft_seconds": null,
"tokens_generated": 256,
"vram_used_gb": 6.99,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-16k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~16K context fill (12020 prompt tokens). Prompt eval 437.1 tok/s. VRAM 7154 MiB. -9.4% from baseline."
} | {
"date": "2026-05-15T00:00:00",
"run_type": "warm_server"
} |
zai-org/GLM-4.7-Flash | glm-4.7-flash | zai-org | Q4_K_M | GGUF | moe_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": false,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": 35,
"ngl"... | {
"tok_per_sec": 25.5,
"ttft_seconds": null,
"tokens_generated": 256,
"vram_used_gb": 6.99,
"ram_used_gb_peak": null,
"prompt_id": "context-fill-24k"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "~24K context fill (18020 prompt tokens). Prompt eval 422.3 tok/s. VRAM 7160 MiB. -14.7% from baseline. Flattest context degradation of all models tested. MLA compresses KV cache."
} | {
"date": "2026-05-15T00:00:00",
"run_type": "warm_server"
} |
zai-org/GLM-4.7-Flash | glm-4.7-flash | zai-org | Q4_K_M | GGUF | moe_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": false,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": 35,
"ngl"... | {
"tok_per_sec": 29.5,
"ttft_seconds": null,
"tokens_generated": 2048,
"vram_used_gb": 6.99,
"ram_used_gb_peak": null,
"prompt_id": "sustained-throughput-2048"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "Sustained throughput test: asked for 2048, got 2048 (full budget). Speed held at 29.5 tok/s. Thinking mode active. Quality: 4/6 pass, 2/6 fail. Code gen fail (reasoning consumed all tokens). Hallucination fail (fabricated complete paper with specific det... | {
"date": "2026-05-15T00:00:00",
"run_type": "warm_server"
} |
zai-org/GLM-4.7-Flash | glm-4.7-flash | zai-org | Q4_K_M | GGUF | moe_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": false,
"thinking_disabled": false,
"temperature": 0,
"top_p": null,
"ncmoe": 35,
"ngl"... | {
"tok_per_sec": 29.9,
"ttft_seconds": null,
"tokens_generated": 786,
"vram_used_gb": 6.99,
"ram_used_gb_peak": null,
"prompt_id": "ncmoe-sweep"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe sweep results: ncmoe40=25.8tok/s(4951MiB), ncmoe35=28.8tok/s(6704MiB, sweet spot), ncmoe32=32.1tok/s(7789MiB, tight), ncmoe30=24.6tok/s(7823MiB, VRAM cliff). Flash attention OFF required for MLA arch."
} | {
"date": "2026-05-15T00:00:00",
"run_type": "warm_server"
} |
openai/gpt-oss-20b | gpt-oss-20b | openai | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 22,
"ttft_seconds": null,
"tokens_generated": 1024,
"vram_used_gb": 2.14,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=30 (all 24 MoE layers' experts on CPU). 21B total, 3.6B active, 32 experts/layer, 4 active. OpenAI Apache 2.0. Native MXFP4 FFN weights. Maximum headroom config (6 GB free). This is the config used for agentic coding (10/10 tasks completed via Pi a... | {
"date": "2026-05-20T00:00:00",
"run_type": "cold_server"
} |
openai/gpt-oss-20b | gpt-oss-20b | openai | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 25.7,
"ttft_seconds": null,
"tokens_generated": 1024,
"vram_used_gb": 3.78,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=20 (4 MoE layers' experts on GPU, 20 on CPU). ~407 MiB per MoE layer on GPU. tok/s estimated from client-side measurement (24.86) with ~3.4% HTTP overhead adjustment."
} | {
"date": "2026-05-20T00:00:00",
"run_type": "cold_server"
} |
openai/gpt-oss-20b | gpt-oss-20b | openai | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 39.45,
"ttft_seconds": null,
"tokens_generated": 1024,
"vram_used_gb": 7.02,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=12 SWEET SPOT at 32K (12 MoE layers' experts on GPU, 12 on CPU). 7024 MiB VRAM with 1164 MiB headroom. Best balance of speed and safety. Architecture: 24 layers, 32 experts/layer, 4 active, n_embd=2880. Per-layer expert cost ~407 MiB on GPU."
} | {
"date": "2026-05-20T00:00:00",
"run_type": "cold_server"
} |
openai/gpt-oss-20b | gpt-oss-20b | openai | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 16384,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 40.6,
"ttft_seconds": null,
"tokens_generated": 1024,
"vram_used_gb": 6.91,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=12 at 16K context. Context scaling nearly flat vs 32K (40.60 vs 39.45 tok/s). KV cache is small relative to expert weights at this model size."
} | {
"date": "2026-05-20T00:00:00",
"run_type": "cold_server"
} |
openai/gpt-oss-20b | gpt-oss-20b | openai | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 32768,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 46.64,
"ttft_seconds": null,
"tokens_generated": 1024,
"vram_used_gb": 7.83,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=10 MAX SPEED (14 MoE layers' experts on GPU, 10 on CPU). 7832 MiB VRAM with only 356 MiB headroom. Fastest config but tight on VRAM. No CUDA cliff observed despite 7.8 GB usage."
} | {
"date": "2026-05-20T00:00:00",
"run_type": "cold_server"
} |
liquid-ai/LFM2-24B-A2B | lfm2-24b-a2b | liquid-ai | Q4_K_M | GGUF | moe_partial_cpu_expert_offload | {
"gpu": "NVIDIA GeForce RTX 4060 Ti",
"vram_gb": 8,
"system_ram_gb": 32,
"ram_spec": "DDR5-6000 CL36",
"cpu": "AMD Ryzen 5 7600X",
"platform": "WSL2 Ubuntu 26.04 on Windows 11"
} | {
"engine": "llama-server (turboquant b9082-5aeb2fdbe)",
"context_length": 16384,
"gpu_offload": "max",
"expert_offload_to_cpu": true,
"kv_cache_quant_k": "q8_0",
"kv_cache_quant_v": "q8_0",
"flash_attention": true,
"thinking_disabled": false,
"temperature": 0.7,
"top_p": null,
"ncmoe": null,
"n... | {
"tok_per_sec": 37.28,
"ttft_seconds": null,
"tokens_generated": 1024,
"vram_used_gb": null,
"ram_used_gb_peak": null,
"prompt_id": "moe-explainer-500w"
} | {
"quality_5pt": null,
"hit_length_target": null,
"notes": "ncmoe=30 (10 layers experts on GPU, 30 on CPU). ngl=999, np=1, threads=6. VRAM projected 4520 MiB / 7063 MiB free (2540 MiB headroom). 64 experts/layer, 4 active. Comfortable but underutilising GPU — lots of headroom remaining."
} | {
"date": "2026-05-13T00:00:00",
"run_type": "cold_server"
} |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.