Home › Evidence › Records › clm-0083
⚠ This notebook has stopped taking notes: newest record is 23 days old (eng-0280), against an expected cadence of 14 days.

clm-0083

measured-heremed ●●○
citable URL: https://halobench.com/records/clm-0083/ — this address never moves; the anchor /records/#clm-0083 keeps resolving

Ling-3.0-flash's corrected post-PR-26608 Q4_K_M GGUF completed a valid full 26-task tau2 airline run on aihydra at 0.500 mean reward (13/26), with 170 tool-call messages, 0 empty assistant messages, 0 infrastructure errors and 0 max-step cuts. The same ROCm/plain-decode config passed the house 4-item guard and measured 36.08 t/s decode at d0, 32.89 t/s at d32768. Energy was 69.218 Wh for the whole tau2 window, or 5.32 Wh per correct answer at 30.3 p/kWh.

verified 2026-08-18 · volatility medium
evidence run-0435 run-0436 run-0437 run-0438 run-0439 eng-0173

Note — the record's own working

METHOD -- Ling-3.0-flash Q4_K_M from bloomer010/Ling-3.0-flash-GGUF, sha256 bcce6e32799749e8e52a52c161127b989db52430e00fbebc77f4d61ef94e754d, on llama.cpp post-BailingMoE3 build 7077abb. ROCm only, f16 KV, --load-mode none, -fa on, --parallel 1, --jinja, temperature 0, max_steps 200, seed 42, pinned user simulator openrouter/anthropic/claude-haiku-4.5. No speculative decoding or MTP path was exercised; the candidate's shipped MTP activation state remains a separate open question. Energy joined same session from raw HA sensor.hardware_ai_hydra_energy history, interpolated to run-meta windows.

Cited by — computed at build time, never stored

model pages ling-30-flash
candidate gate history ling-30-flash