Huihui Qwen3.8 qualification addendum — 2026-09-19

Status: qualification evidence only. Operator authorization, applied state, and client distribution are private.

The earlier 8K scout remains historical rejected/no-promotion evidence for its original runtime. The compatible-runtime follow-up remains the evidence basis: functional gates passed, matched 8K MTP3/no-spec capacity cells completed 12/12, and the bounded 32K MTP3 cell completed 6/6. Failed 128-word warmups, boundary normalization, the unpaired 32K comparison, and higher GPU use than the cross-profile GGUF control remain retained limitations.

Fresh baked-image gate

The baked MTP3 32K image passed a fresh quality gate at the nominal 31K target. Tool calling, unified-diff editing, timeout triage, and session recall passed 3/3. Fresh vision passed 12/12 and baked capacity passed 18/18. The gate summary is sanitized.

Historical benchmark publication staging

Six eligible capacity artifacts are retained: no-spec 8K n=12, MTP3 8K n=12, GGUF-control 8K n=12, MTP3 32K n=6, and final baked short n=12 and long n=6. Failed 128-word warmups remain excluded from performance import but retained as failures. Public records do not identify operational dashboards, assignment, readiness, or distribution.

Limits

Browser acceptance remains unverified. Pinned build inputs are retained, but a fresh-host identical digest is not claimed. Operator records and recovery selection are private.

64K C1 reference

The 64K qualification reference passed direct preflight 7/7, routed preflight 6/6, vision 12/12, and native Hermes 9/9. Descriptive short capacity passed 12/12 at 182.1 mean decode tokens/s; long capacity passed 6/6 at 167.1 with 60,769-60,776 actual prompt tokens. See the extension finding.