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

Now

Generated 2026-08-30 by scripts/generate-now.mjs from structured candidate and claim records. Do not edit this page by hand; run npm run generate:now after the record moves.

Running now

  • The latest public-record campaign is complete. Qwen3.8-Flash-Next full-STRIX completed with a bounded/partial evidence package. Operational scheduling remains private; this page reports only independently reviewed evidence in the public record.

Latest measured outcomes

  • Qwen3.8-Flash-Next full-STRIX ROCmFP4 — bounded/partial. Standard Tau2 attempted all 26 tasks: 19/24 evaluated passes, with two infrastructure errors retained separately rather than zero-imputed. Exactly two original calls reached the inherited 8,192-token output ceiling; a later two-task 16k recovery is held for missing lossless transport evidence and contributes no mixed-budget score. Corrected thinking-off QSA prompt processing, selected nonzero-output generation rows and separate prefix-reuse traces are admitted. A separate post-hoc wall-meter join recovers the retained full-run UTC window without claiming campaign-time metering compliance. Incomplete PLE mincore, engine scatter, zero-byte generation rows and the full-window HTTP 400 remain explicit exclusions (clm-0127).
  • DeepSeek-V4-Flash-0731 — benched. The carried Vulkan fork passed the house guard and completed the bounded f16-KV series through d262144 without device loss. This supersedes the stock-build-only device-loss conclusion, but does not establish general Vulkan safety because the payload bundles multiple changes and did not run the full capability suite (clm-0084).
  • Gemma 4 26B-A4B — rejected. HG-001 stopped at the declared guard gate: 3/4, with needle retrieval lost at 2048 tokens. No smoke, throughput, Vulkan or energy result exists; the evidence supports a retrieval-envelope/probe failure, not a runtime crash or broad quality verdict (clm-0085).
  • Ling-3.0-flash — benched. HO-002 proved that the shipped NextN head activates with nonzero acceptance, but n_max 1–3 produced matched decode ratios of 1.007x, 0.968x and 0.850x, all below the predeclared 1.10x threshold. Plain decode remains the recommendation (clm-0086).

Current decisions

  • Keep the KingJones full-STRIX ROCmFP4 campaign separate from the historical Unsloth UD-Q4_K_XL/Vulkan series; their artifact, runtime and backend identities do not transfer. The new record is bench evidence, not a production recommendation or publisher-equivalent result.
  • DeepSeek’s carried Vulkan result is promising fork-specific evidence, not attribution to a single patch and not an acquittal of stock Vulkan.
  • Gemma 4 26B-A4B does not advance from its measured rejection without a new, probe-specific hypothesis.
  • Ling MTP activation is closed for this bounded workload; supported does not mean faster.
  • Candidate gate histories in content/candidates/*.yaml and the generated candidate board are canonical. docs/candidates.md records the editorial policy and reconciliation notes, not a second status table.