cfg-0001
backfilled
citable URL: https://halobench.com/records/cfg-0001/ — this address never moves; the anchor /records/#cfg-0001 keeps resolving
| model | unsloth/Qwen3.6-35B-A3B-GGUF · UD-Q4_K_XL |
| engine | ggml-org/llama.cpp unknown-2026-06 · rocm · host aibeast (igpu) |
| flags | -ngl 999 --no-mmap --parallel 1 -fa 1 --jinja -c 131072 (reasoning OFF / enable_thinking:false; q8_0 KV; no MTP — this GGUF ships no heads) |
| template | not recorded at test time |
| tree | upstream — stock |
aged evidence — reconstructed from the archive. llama.cpp commit, chat-template hash and the memory block were not recorded at test time. Quant, backend, context, KV quant, serving flags and the reasoning setting are from the Phase 3C write-up.
Capability basis
unmeasured — performance numbers on this config stand on a guard alone, not an established capability
production node(s): qwen36-35b-a3b
Runs on this config (3)
| run | date | kind | suite | metrics |
|---|---|---|---|---|
| run-0001 | 2026-06-27 | performance | tool-loop@v2.1 | decode_tps=69 · prefill_tps=354 · ttft_ms=180 |
| run-0003 | 2026-06-28 | capability | conceptual-c1-c8-blind@v1 | time_to_correct_answer_s=1.7 |
| run-0004 | 2026-06-27 | guard | tool-loop@v2.1 | tool_call_success=0.94 · turns_before_degradation=10+ · empty_arg_calls=0 |