cfg-0160
backfilledfork
citable URL: https://halobench.com/records/cfg-0160/ — this address never moves; the anchor /records/#cfg-0160 keeps resolving
| model | Qwen3.6-35B-A3B-MTP-UD-Q4_K_M.gguf · UD-Q4_K_M · rev 0b21525e972670ed59e1812e170b27c26355381f0656ecc4e25617ece7dac58b |
| engine | ggml-org/llama.cpp 2586f6eddae19bef3dd21a1a0b109cca7bcf1c32 · Vulkan0/Strix-Halo (strix-halo-v0610 fork via RADV, NOT ggml-org ROCm0) · host aihydra (igpu) |
| flags | HO-013 cell1 n4-off (MTP n_max=4, reasoning OFF) capability fingerprint on the v0.6.10 fork/Vulkan runtime 2586f6edd. Mirrors cfg-0150 (f16/f16 KV, -ngl 999 -fa 1 -c 32768 --load-mode none --parallel 1 --jinja -rea off -n 4096 --slots -ctk f16 -ctv f16 -b 2048 -ub 512 -t 16, temp 0 seed 42 max_tokens 4096) but --spec-draft-n-max 4 (--spec-type draft-mtp -ctkd f16 -ctvd f16; draft KV stays f16 = no KV quant). n_max=4 ONLY; no reasoning-on, no KV quant, no alt drafter. |
| template | not recorded at test time |
| tree | fork — carrying con-0002, con-0009 |
aged evidence — reconstructed from the archive. Recorded from raw cell1-n4-off/fingerprint.txt (Vulkan/RADV build 2586f6edd, model sha 0b21525e, spec=n4 reasoning=off, base_flags + spec_flags recorded verbatim) and gates.txt G3/G6. ADMITTED full-26 capability arm (17/26, mean 0.653846) via HO-013-REVIEW. Scope strict to this single model/quant/build, MTP n2/n4 + reasoning on/off matrix only; no cross-model/KV-quant/DFlash2/n>4 claim.
Capability basis
measured on this config
Runs on this config (3)
| run | date | kind | suite | metrics |
|---|---|---|---|---|
| run-0597 | 2026-08-23 | guard | eos-cliff-c32768-varied-fullprompt | tasks_passed=8 · tasks_total=8 |
| run-0598 | 2026-08-23 | guard | guard-c32768-depth8000 | tasks_passed=4 · tasks_total=4 |
| run-0599 | 2026-08-23 | capability | tau2-bench-airline@668d3bc | tasks_passed=17 · tasks_total=26 · mean_score=0.6538461538461539 |
Cited by — computed at build time, never stored
claims clm-0113