Skip to content

Qwen3.6-27B NVFP4+MTP evaluation baseline

Point-in-time record, 2026-07-12. This run restored the previously tested sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP candidate on Primary Node's RTX PRO 6000, ran the existing built-in bakeoff as a baseline, and then ran the newer session-derived --suite-file evaluation unchanged. The built-in bakeoff passed. The session-derived suite reported 0/5, but that score remains invalid for ranking or promotion under the known-broken cross-model protocol.

Configuration

Field Tested value
Served model qwen36-27b-nvfp4-mtp
Checkpoint sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP
Host Primary Node, Windows 11, Docker Desktop/WSL2
GPU RTX PRO 6000 Blackwell Max-Q, 96 GB, sm_120
Engine vLLM 0.23.1rc1.dev531+ga65f93fb2
Quantization ModelOpt NVFP4, text-only, BF16 MTP head
KV cache FP8
Context / sequences 262,144 / 2
Speculative decoding Qwen MTP, 3 speculative tokens
Endpoint http://127.0.0.1:39027/v1
Managed serve cand-qwen36-heavy-mtp

The managed swap stopped the unused heavy / vllm-gptoss120 serve before starting this candidate. The checkpoint was 18.29 GiB; target plus draft model loading took 158.13 seconds and the healthy post-start GPU reading was 88,112 MiB. vLLM selected the FlashInfer CUTLASS NVFP4 linear kernel, the Triton/FLA GDN prefill kernel, and a 1,600-token hybrid cache block. The engine warned that the checkpoint does not provide calibrated FP8 KV q/prob scaling factors and used 1.0; retain that as an accuracy caveat for later context work.

Correctness gate

Preflight ran with thinking disabled, a 128,000-token needle, and 20 shared-prefix tool calls. All checks passed: short coding in 5.3 seconds, structured JSON, the 128K needle in 25.6 seconds, and 20/20 valid tool calls.

Existing built-in baseline

Raw artifact: current-built-in-eval.json.

Check Result
Context pass; 128,540 prompt tokens, 35.01 s TTFT, 35.70 s end-to-end
Tool call pass
Multi-turn session recall pass
Unified-diff intelligence pass
Parallel timeout triage pass

The built-in intelligence result was 2/2 and the complete built-in bakeoff had no failures. This is the requested current-eval baseline. It is a single run, not repeated quality evidence, and the context measurement had prefix caching disabled by the serve recipe.

New session-derived deterministic suite

Suite source: planning-milestone-execution.suite.json. Raw result: new-session-derived-eval.json.

Eval Result Missed literal checks
Low-overhead dashboard architecture fail stdlib, bounded retention, external evidence, degraded capability
Milestone dependency order fail exact ordered chain
Proof-buffer recovery fail reject, capture-evidence, resubmit/strict
Resumable CI monitor fail persist/resume and existing pull-request reuse
Local/remote main reconciliation fail explicit preservation of the untracked file

Overall: 0/5 passed. Unlike the GPT-OSS budget-starvation control, every case produced substantive visible answer content and Qwen thinking was disabled through its supported chat-template control. This isolates a different protocol problem: one-shot exact-substring contracts reject directionally reasonable answers that do not reproduce the requested operational vocabulary.

The result still cannot be used as a cross-model quality score. The suite's context_bucket remains metadata rather than supplied context, the artifact retains only excerpts rather than full output and finish-reason/reasoning-channel evidence, and the run was not repeated. No routing profile, serve recommendation, or promotion changed.

Baseline conclusion

The checkpoint remains operationally healthy at its native 262K serve window and passes the current built-in correctness/intelligence path at 131K. The new suite's 0/5 is useful as a baseline for repairing the evaluation protocol, not as evidence that the model has zero planning quality.