Skip to content

Candidate intake

Observed 2026-09-19. The producer's model card and build manifest are external priors, not local qualification.

  • Model: lyf/Qwen3.8-27B-Huihui-Abliterated-NInfer-NVFP4
  • Revision: 181446902fc777c479749e98cf2abf2250263a8d
  • Source revision: d42ca8978c5a66e92c3446d46e8adfe03ef692ff
  • NInfer: a99407c63fc5bbd25d9fb597cbb8ab352bdb01ef
  • Artifact bytes: 21,492,695,040; identical size to prior NInfer artifact.
  • Artifact SHA256: f21f308d3b23ccd627071cd015e413db08deee4356643900518e2b251750fdc2
  • Producer observed date: 2026-08-19 (31 days old).
  • Hardware relevance: exact RTX 5090 product; different host and driver.
  • Decision impact: bounded 8K scout before capability/context expansion.
  • Producer's 204,800 context claim covers startup/API smoke, not long-context quality.

Pinned source

The September 3 local NInfer campaign supplies a historical comparison, not a matched current baseline. It used another artifact/runtime and 252,928 context. Fresh matched cells are required before claiming memory or speed improvement.

The deployed infrastructure already has registry-based benchmark import and Workbench card generation; an older local checkout still has a hard-coded allowlist. Use the deployed source when preparing the repeatable update.