Skip to content

Publication summary: Huihui Qwen3.8 64K C1

This copy derives from the linked native artifacts and dated finding.

Canonical facts

  • Model: lyf/Qwen3.8-27B-Huihui-Abliterated-NInfer-NVFP4@181446902fc777c479749e98cf2abf2250263a8d; served qwen38-huihui-ninfer-schemafix-mtp3-64k.
  • Runtime: NInfer 70434721b1ae29d0616f3de9b376c8a4d91590b5; image sha256:44377d56374ed4b832c2ca6e9265da492cf2ad79a297008bae92f442d27edfb6.
  • Setup: one RTX 5090, Windows/WSL2, NVFP4/mixed FP8 weights, INT8 group-64 KV, MTP3, no thinking, vision, 65,536 context tokens, C1.
  • Recipe: exact sanitized managed recipe, bound to qualification identity.
  • Measurement path: direct online capacity with unique prompts; native routed protocol and Hermes checks in bounded compatibility probes. No kernel-only or matched 64K no-spec measurement.
  • Headline: descriptive, canary-free C1 mean decode rates of 182.1 tokens/s for 12 short requests and 167.1 for six long requests. These are not finalist comparative performance claims.
  • Capabilities: direct preflight 7/7, routed preflight 6/6, repeated bounded quality, vision 12/12, Hermes 9/9; quality request had 62,455 actual prompt tokens.
  • Caveats: post-workload memory 24,224 MiB is higher than the historical GGUF control; no endurance, transient peak or interactive browser acceptance claim.
  • Decision: qualified 64K reference; actual deployment and recovery state is private.
  • Canonical evidence: dated finding.
  • Artifact set: manifest, index.

Claim ledger

Claim Conditions Evidence
Bounded protocol and quality passed Pinned 64K C1 profile, repeated core/session and one long context request preflight, quality, routed preflight
Vision 12/12 Hashed synthetic image corpus, two repetitions per case vision
Descriptive decode 182.1 / 167.1 tokens/s C1, unique prompts, 32 exact words, n12/n6, no canaries short, long
Hermes reached the local secondary Native anvil provider, exact alias/provider, tool-result continuation native client receipt
GPU use 24,224 MiB After retained workloads, not a transient peak memory
Oversized default preflight failed Mistaken 128K/C20 defaults, then corrected to in-range C1 failure, friction

No social post was requested or sent.