8GB VRAM Local LLMs - Practitioner Tested
Collection
11 models on RTX 4060 Ti 8GB. speed king: Llama 1B (228 tok/s). quality king: Gemma 4 E4B (6/6). agentic king: gpt-oss-20b (10/10). • 21 items • Updated • 5
config stringclasses 7
values | runtime stringclasses 2
values | kv_cache_k stringclasses 2
values | kv_cache_v stringclasses 5
values | context_k int64 32.8k 65.5k | np_slots int64 1 4 | graph_splits int64 62 82 | kv_vram_mib int64 340 1.36k | checkpoint_mode stringclasses 2
values | agent stringclasses 3
values | system_prompt_tokens int64 1.9k 13.5k ⌀ | context_depth_k float64 3.5 50 | decode_tok_s float64 9 35.8 | prompt_eval_tok_s float64 88 461 ⌀ | test_type stringclasses 3
values | notes stringlengths 11 71 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
baseline_hermes | llama.cpp | q8_0 | q8_0 | 32,768 | 4 | 62 | 1,360 | default | hermes | 13,500 | 3.5 | 31.4 | null | hermes_round | original smoke test config |
phase1_np1 | llama.cpp | q8_0 | q8_0 | 32,768 | 1 | 62 | 340 | default | none | null | 3.5 | 34.94 | null | curl | added -np 1 |
phase2_q4v_64k | llama.cpp | q8_0 | q4_0 | 65,536 | 1 | 82 | 520 | default | none | null | 3.5 | 31.22 | null | curl | V cache q4_0 at 64K |
phase2_q4v_64k | llama.cpp | q8_0 | q4_0 | 65,536 | 1 | 82 | 520 | default | hermes | 13,500 | 14 | 9 | null | hermes_round | 82 splits killed Hermes speed |
phase2_q4v_48k | llama.cpp | q8_0 | q4_0 | 49,152 | 1 | 82 | 390 | default | none | null | 3.5 | 30.52 | null | curl | 48K still has 82 splits |
turboquant_turbo4 | turboquant | q8_0_auto | turbo4 | 65,536 | 1 | 62 | 510 | default | none | null | 3.5 | 13.93 | null | curl | auto-asymmetric upgraded K to q8_0; turbo4 decompression too aggressive |
turboquant_turbo2 | turboquant | q8_0 | turbo2 | 65,536 | 1 | 62 | 510 | default | none | null | 3.5 | 35.13 | null | curl | turbo2 keeps 62 splits at 64K |
turboquant_turbo2 | turboquant | q8_0 | turbo2 | 65,536 | 1 | 62 | 510 | default | hermes | 13,500 | 14.3 | 33.59 | null | hermes_round | confirmed speed holds in Hermes |
turboquant_turbo2 | turboquant | q8_0 | turbo2 | 65,536 | 1 | 62 | 510 | default | hermes | 13,500 | 14.3 | 35.76 | null | hermes_round | second call same session |
turboquant_turbo3 | turboquant | q8_0 | turbo3 | 65,536 | 1 | 62 | 510 | default | none | null | 3.5 | 35.17 | null | curl | turbo3 matches turbo2 speed with better compression |
turboquant_turbo3 | turboquant | q8_0 | turbo3 | 65,536 | 1 | 62 | 510 | default | pi | 1,900 | 3.5 | 34.18 | null | pi_round | Pi system prompt 7x smaller than Hermes |
turboquant_turbo3 | turboquant | q8_0 | turbo3 | 65,536 | 1 | 62 | 510 | default | pi | 1,900 | 17.2 | 31.17 | null | pi_round | speed holds at 17K context |
turboquant_turbo3 | turboquant | q8_0 | turbo3 | 65,536 | 1 | 62 | 510 | default | pi | 1,900 | 26.5 | 14.64 | 88 | pi_round | VRAM cliff from checkpoint accumulation |
turboquant_turbo3 | turboquant | q8_0 | turbo3 | 65,536 | 1 | 62 | 510 | disabled | none | null | 3.5 | 34.25 | 370 | curl | checkpoints disabled via --checkpoint-every-n-tokens -1 |
turboquant_turbo3 | turboquant | q8_0 | turbo3 | 65,536 | 1 | 62 | 510 | disabled | none | null | 20 | 31.54 | 461 | curl | no cliff at 20K |
turboquant_turbo3 | turboquant | q8_0 | turbo3 | 65,536 | 1 | 62 | 510 | disabled | none | null | 26 | 30.26 | 443 | curl | no cliff at 26K — was 14.64 with checkpoints |
turboquant_turbo3 | turboquant | q8_0 | turbo3 | 65,536 | 1 | 62 | 510 | disabled | none | null | 35 | 29.62 | 455 | curl | smooth degradation continues |
turboquant_turbo3 | turboquant | q8_0 | turbo3 | 65,536 | 1 | 62 | 510 | disabled | none | null | 50 | 27.71 | 450 | curl | 50K context at 28 tok/s — no cliff |
RTX 4060 Ti 8GB — turboquant KV cache benchmark (Qwen3.6-35B-A3B)
practitioner-tested benchmarks of turboquant KV cache types vs standard llama.cpp on an RTX 4060 Ti 8GB with 32GB DDR5-6000 RAM.
hardware
| component | spec |
|---|---|
| GPU | NVIDIA RTX 4060 Ti, 8 GB VRAM |
| CPU | AMD Ryzen 5 7600X (6c/12t) |
| RAM | 32 GB DDR5-6000 dual-channel |
| OS | Windows 11 + WSL2 Ubuntu 26.04 |
model
Qwen3.6-35B-A3B-UD-Q4_K_M (22.1 GB). hybrid SSM+attention architecture — 10/40 layers use attention, 30/40 use Gated Delta Net (recurrent). MoE with partial offload: -ncmoe 30 (10 layers experts on GPU, 30 on CPU).
what this dataset covers
turboquant KV cache types — turbo2, turbo3, turbo4 vs standard q4_0 and q8_0. turboquant is a llama.cpp fork (github.com/TheTom/llama-cpp-turboquant) that adds aggressive KV cache compression with auto-asymmetric behaviour (K stays q8_0 for high GQA ratios, only V is compressed).
graph splits — the hidden speed determinant on 8GB VRAM. standard llama.cpp V q4_0 at 64K context produces 82 graph splits (GPU-CPU transfers per step) vs 62 at 32K. turboquant turbo2/turbo3 keep 62 splits at 64K. the split count matters more than raw VRAM usage at this boundary.
context checkpoint impact — llama.cpp creates context checkpoints every 8192 tokens during prefill (~63 MiB each in VRAM). at 26K context, 5 checkpoints = 315 MiB, pushing past the 7 GB VRAM cliff. disabling with --checkpoint-every-n-tokens -1 eliminates the cliff entirely. this is a novel finding — I have not seen it documented elsewhere.
agent harness overhead — Hermes Agent v0.13.0 (
13.5K token system prompt) vs Pi coding agent v0.74.0 (1.9K tokens). same decode speed, 7x smaller system prompt = 13K more usable context.speed curve at depth — decode speed from 3.5K to 50K context with checkpoints disabled. smooth degradation (~2 tok/s per 10K tokens) from attention scaling on the 10 attention layers. no cliff.
key findings
- turbo3 is the sweet spot for 8GB VRAM. same speed as turbo2 (35 tok/s) with better V cache compression. turbo4 is too aggressive (13.93 tok/s — decompression overhead kills it).
- graph splits > raw VRAM as the speed determinant at this boundary. 62 splits = fast, 82 splits = broken.
- context checkpoints are a hidden VRAM eater. disabling them is free on Qwen3.6 (hybrid architecture invalidates them anyway).
- final config: 64K context, 28-34 tok/s across all depths, no cliff.
recommended config
-ngl 999 -ncmoe 30 -c 65536 -np 1 -fa on --cache-type-k q8_0 --cache-type-v turbo3 --no-cache-prompt --checkpoint-every-n-tokens -1
methodology
- all tests on same hardware, same Q4_K_M quant, same -ncmoe 30 partial offload
- curl tests: single 300-token completion request (warm run)
- agent tests: real multi-step tasks via Hermes or Pi
- decode speed from llama-server slot print_timing logs
- tested 2026-05-11 and 2026-05-12
related datasets
- witcheer/windows-rtx-4060ti-8gb-bench-2026-05 — dense model benchmarks
- witcheer/windows-rtx-4060ti-8gb-moe-offload-bench-2026-05 — MoE offload benchmarks
collection
- Downloads last month
- 20