clm-0081
measured-herehigh ●●●
citable URL: https://halobench.com/records/clm-0081/ — this address never moves; the anchor /records/#clm-0081 keeps resolving
Qwen3-Coder-Next 80B ships NO speculative-decode or MTP head in its official Q8_0 GGUF. This job's own from-scratch GGUF header parser (gguf_probe2.py, no numpy/gguf-py dependency) read all 807 tensors across all 4 shards directly and ran an exhaustive needle scan for nextn/mtp/eagle/medusa/draft on every tensor name: ZERO hits on every needle, on every shard. The base model's own config.json independently corroborates this before any download: standard Qwen3NextForCausalLM causal-LM config, no nextn/draft-head keys of any kind. Plain decode throughout this bench -- no spec-decode arm was attempted or is possible against this artifact.
verified 2026-08-17 · volatility low
Note — the record's own working
Checked BEFORE claiming anything MTP-related, per this lab's own standing caution (burned once by a mislabeled non-MTP artifact): the acquired-gate candidate record (content/candidates/qwen3-coder-next-80b.yaml) ran this exact scan on 2026-08-17 immediately post-download, across all 4 shards independently (not just the first), before any bench work began. Matches the same house-standard verification ornith-35b (clm-0070) and qwen36-27b-mtp's candidate record both carry. No EOS-cliff watch applies here (that hazard class is specific to models WITH an MTP head run past spec-draft-n-max >= 4; this candidate has no MTP head to watch).
Cited by — computed at build time, never stored
model pages qwen3-coder-next-80b
candidate gate history qwen3-coder-next-80b