| { |
| "schema_version": 1, |
| "created_utc": "2026-07-28T02:36:27.566499+00:00", |
| "status": "adaptive_release_candidate_qualified", |
| "scope": "single-machine batch-1 Qwen3.6 output-projection offload", |
| "model": { |
| "name": "Qwen3.6-35B-A3B Q5_K_M", |
| "base_model": "Qwen/Qwen3.6-35B-A3B", |
| "gguf_source": "bartowski/Qwen_Qwen3.6-35B-A3B-GGUF", |
| "gguf_sha256": "EC438D1326EB11D8BB94785C54739CD4B60B01AE32BD8532E232A046B19D296D" |
| }, |
| "platform": { |
| "os": "Windows 11 ARM64", |
| "soc_class": "Snapdragon X2 Elite Extreme", |
| "transformer_accelerator": "Adreno X2-90 OpenCL GPU", |
| "output_projection_accelerator": "Hexagon v81 NPU via QNNExecutionProvider", |
| "power": "AC connected; battery status 2; 95-97%; ASUS Performance mode" |
| }, |
| "runtime": { |
| "strategy": "full W8A16 NPU projection with conditional W16A16 shard refinement", |
| "primary_shards": 2, |
| "refinement_shards": 8, |
| "refine_margin": 0.15, |
| "cpu_fallback_allowed_for_npu_graphs": false, |
| "libllama_sha256": "D15D0734E9D1718E6D3D85F02B06DF7E39CA72B0EF8284ED591F9F0611FFBDF9" |
| }, |
| "release_candidate": { |
| "tokens_per_second": 24.257476, |
| "samples_tokens_per_second": [24.1019, 24.2887, 24.3818], |
| "gain_vs_faster_trailing_control_percent": 2.8240472907580028, |
| "gain_vs_two_control_mean_percent": 7.851746040687679, |
| "component_quality": { |
| "samples": 33, |
| "global_top1_matches": 33, |
| "global_top1_match_rate": 1.0, |
| "top5_overlap_mean": 1.0, |
| "top10_overlap_mean": 0.9878787878787879, |
| "cosine": 0.9999566447246216, |
| "nrmse": 0.00933828799354565, |
| "rmse": 0.019982964555496554, |
| "mae": 0.015304419790575721, |
| "max_abs": 0.2505837678909302, |
| "kl_divergence_mean": 0.00019464064398776732, |
| "kl_divergence_max": 0.0005660149465664289 |
| }, |
| "component_latency_ms": { |
| "warmup": 5, |
| "iterations": 100, |
| "median": 6.8199000000000005, |
| "p95": 9.73378, |
| "mean": 7.425287, |
| "minimum": 6.4286, |
| "maximum": 12.72 |
| }, |
| "behavioral_parity": { |
| "status": "exact_text_parity", |
| "exact_text_match": true, |
| "generated_tokens": 192 |
| }, |
| "live_execution": { |
| "refined_decode_calls": 59, |
| "total_decode_calls": 193, |
| "refined_parts_total": 64, |
| "npu_head_median_ms": 4.551 |
| } |
| }, |
| "alternatives": { |
| "static_split_quality": { |
| "tokens_per_second": 23.81049, |
| "global_top1_matches": 33, |
| "samples": 33, |
| "cosine": 0.9999746411320438, |
| "nrmse": 0.007150120560477397, |
| "status": "component_gate_passed" |
| }, |
| "full_w16_maximum_precision": { |
| "live_benchmark_pending": true, |
| "global_top1_matches": 33, |
| "samples": 33, |
| "cosine": 0.9999999659656381, |
| "nrmse": 0.000262637965804305, |
| "kl_divergence_mean": 1.0580494337893319e-7, |
| "status": "component_gate_passed" |
| }, |
| "full_w8_performance_only": { |
| "tokens_per_second": 25.580063, |
| "global_top1_matches": 32, |
| "samples": 33, |
| "status": "rejected", |
| "reason": "one captured near-tie changed top-1 under W8" |
| } |
| }, |
| "controls": { |
| "control_a_tokens_per_second": 21.391753, |
| "control_b_tokens_per_second": 23.591248, |
| "two_control_mean_tokens_per_second": 22.4915005, |
| "warning": "control drift was material; use the faster trailing control for the conservative comparison" |
| }, |
| "limitations": [ |
| "one physical machine", |
| "33 calibration states", |
| "one 192-token deterministic behavioral comparison", |
| "no task-level UMoX benchmark suite yet", |
| "no energy or long-duration thermal measurement in final qualification", |
| "only the output projection is offloaded to NPU" |
| ] |
| } |
|
|
|
|
