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; servedqwen38-huihui-ninfer-schemafix-mtp3-64k. - Runtime: NInfer
70434721b1ae29d0616f3de9b376c8a4d91590b5; imagesha256:44377d56374ed4b832c2ca6e9265da492cf2ad79a297008bae92f442d27edfb6. - 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.