Skip to content

Qwen3.6 protocol-v2 comparison on RTX PRO 6000

Point-in-time record, 2026-07-12. Four Qwen3.6-27B checkpoints were tested one at a time on Primary Node's single RTX PRO 6000 Blackwell 96 GB. This round used the repaired deterministic protocol-v2 suite, audited a suspicious completion-budget failure, added and tested Unsloth's new NVFP4 checkpoint, and restored ThinkingCap as the resident Heavy quality challenger at http://127.0.0.1:39031/v1. No router profile or production tier was promoted.

Raw artifacts are under 2026-07-12-qwen36-protocol-v2-evidence/. Their committed byte identities are listed in SHA256SUMS.

Tested checkpoints and serving recipes

Candidate Pinned revision Engine / quantization Context / admission
sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP 6f194695406a3bc88a00573187d5b2eecf984a99 vLLM 0.23.1rc1.dev531, ModelOpt NVFP4, FP8 KV, MTP 3 262,144 / five; exact-card calibration used two
Qwen/Qwen3.6-27B-FP8 e89b16ebf1988b3d6befa7de50abc2d76f26eb09 vLLM 0.23.1rc1.dev531, FP8, FP8 KV, MTP 3 262,144 / five
bottlecapai/ThinkingCap-Qwen3.6-27B-FP8 e48255afd77b403446332be0f595868337b36591 vLLM 0.23.1rc1.dev531, FP8, FP8 KV, MTP 3 262,144 / five
unsloth/Qwen3.6-27B-NVFP4 ccdaab7e68af2409599b8949a8f2685703c9bae5 vLLM 0.25.0, compressed-tensors NVFP4, FP8 KV, MTP 2 262,144 / five

The three original candidates pin image digest sha256:907377dddef392f6b679d9c071e1c33c3935b4dc993b61d0352e391a5319ff3e. Unsloth pins amd64 vLLM 0.25.0 manifest digest sha256:e1c1ff1af9a15921bfa11d1d95047258c1797392cdbfa296e7639da446b23f97. Operator inspection of the latter image reported flashinfer-python==0.6.13 and nvidia-cutlass-dsl==4.5.2, matching Unsloth's published minimum recipe. Operator-observed runtime logs selected FlashInferCutlassNvFp4LinearKernel; those inspection/log transcripts were not retained, so the pinned recipe is reproducible but these runtime facts are not raw-artifact-backed in this change.

Operator-observed preflight passed smoke, structured JSON, the approximately 128K needle, and 20/20 shared-prefix tool calls with thinking disabled for every candidate. Those preflight transcripts were not retained in this change, so these deployment facts are not independently recoverable from the staged raw directory. The prior variation round already recorded 5/5 concurrency completion for the first three. Unsloth's retained raw artifact also records 5/5 at concurrency five in this round.

Matched 1,024-headroom repeated baseline

Each item received 256 visible-answer tokens plus 1,024 reasoning-headroom tokens as one 1,280-token API completion cap. Each item ran three times and needed at least two passes. These are identical fixtures and validators across all four checkpoints.

Candidate ARC stable / attempts ARC wall MMLU-Pro stable / attempts MMLU wall
Community NVFP4+MTP 3/5; 9/15 154.17 s 0/10; 0/30 350.33 s
Official FP8 2/5; 6/15 195.05 s 1/10; 3/30 419.18 s
ThinkingCap FP8 5/5; 15/15 104.36 s 7/10; 21/30 300.24 s
Unsloth NVFP4 4/5; 12/15 152.63 s 1/10; 3/30 418.81 s
Nemotron 3 Super reference 5/5; 15/15 69.91 s 8/10; 23/30 260.91 s

Matched raw runs: community NVFP4 ARC / MMLU-Pro, official FP8 ARC / MMLU-Pro, ThinkingCap ARC / MMLU-Pro, and Unsloth ARC / MMLU-Pro.

Why the low Qwen scores are not intelligence scores

An independent recipe audit found that every community-NVFP4 MMLU attempt consumed exactly the 1,280-token cap and ended with finish_reason=length. Twenty-four of 30 attempts never left the reasoning channel; the other six started a visible answer but were cut off before the final marker. ARC lost six attempts to the same failure class. The reasoning parser, chat-template control, checkpoint revision, and MTP path were all working. Runtime MTP acceptance was approximately 67–83 percent, so speculative decoding was not silently disabled.

The serving recipe did have two audit-worthy differences from the checkpoint's RTX PRO 6000 card: five admitted sequences and 0.92 GPU utilization instead of two and 0.90. The exact-card A/B started cleanly and passed preflight, but those admission settings cannot explain a serial request exhausting its response budget. FP8-KV startup also warned that missing attention scales fall back to 1.0; that remains a quality-confound follow-up, not proof of damage.

The valid conclusion is therefore completion-budget starvation, not model incompetence. The 1,024-headroom rows remain useful as a constrained-efficiency comparison only.

Calibrated operating points

Calibration used one attempt per item to select a budget. Only ThinkingCap's selected 4K point was then confirmed with all three repetitions.

Candidate Selected calibration ARC MMLU-Pro Decision
Community NVFP4+MTP 8,192 5/5 one pass 8/10 one pass; two truncations Did not challenge ThinkingCap; no repeated confirmation
Official FP8 4,096 5/5 one pass 8/10 one pass; two truncations Did not challenge ThinkingCap; no repeated confirmation
ThinkingCap FP8 4,096 5/5 and 15/15 already at 1,024 9/10 stable; 27/30 attempts, 458.78 s Best confirmed quality-slice score
Unsloth NVFP4 8,192 not rerun; 4/5 stable at 1,024 9/10 one pass; one truncation, 371.61 s Same calibration score at twice ThinkingCap's budget; no repeated confirmation

Selected calibrated artifacts: community NVFP4 ARC 8K / MMLU-Pro 8K, official FP8 ARC 4K / MMLU-Pro 4K, ThinkingCap repeated MMLU-Pro 4K, and Unsloth MMLU-Pro 8K.

ThinkingCap's remaining computer-science item still exhausted 4K reasoning in all three confirmation attempts. At 8K it completed but answered incorrectly in the calibration. More budget is therefore not automatically more quality.

Five-session capacity

All candidates completed five of five independent 8K-context requests with thinking disabled. These short-output runs validate admission and scheduling, not long-context five-session residency at five full 262K windows.

Candidate Completed TTFT p50 E2E p50 Aggregate output throughput
Community NVFP4+MTP 5/5 3.22 s 3.75 s 15.74 tok/s
Official FP8 5/5 5.68 s 6.31 s 8.31 tok/s
ThinkingCap FP8 5/5 4.66 s 5.22 s 7.92 tok/s
Unsloth NVFP4 5/5 3.68 s 4.21 s 15.21 tok/s

Unsloth raw capacity runs: concurrency one and concurrency five.

Unsloth's first in-process cold pull was launched without HF_TOKEN loaded and later stalled; the subsequent authenticated dedicated pull completed. That observation does not prove the missing token caused the stall. The canonical models pull command now forwards HF_TOKEN by default, prefers an exported value, falls back to ~/.env, and supports explicit --no-token. The authenticated resumable pull completed all 20 files; cached startup reached health in 239 seconds. ThinkingCap also incurred a roughly five-minute cold weight-load and torch.compile startup when restored. These are operational caveats, not steady-state request latency.

Source age and decision use

Source Observed / source date Age class Evidence type Decision impact
Qwen3.6 official FP8 card observed 2026-07-12; updated 2026-04-24 current official primary native context, model controls, official baseline
Community NVFP4+MTP card observed 2026-07-12; updated 2026-04-29 current community checkpoint recipe exact-card A/B and MTP configuration
ThinkingCap card observed 2026-07-12; updated 2026-07-10 current fine-tune publisher selected fine-tune candidate
Unsloth Reddit announcement published 2026-07-10; observed 2026-07-12 current publisher announcement / community discussion discovered checkpoint and performance claim; local tests required
Unsloth NVFP4 card and recipe observed 2026-07-12; revision updated 2026-07-12 current publisher primary selected engine/dependency/MTP recipe

The Reddit throughput claim is an external prior, not a local result. Its cited B200/128-concurrency setup is not comparable to this serial RTX PRO 6000 quality suite. Local preflight, quality, and five-session artifacts are the decision evidence. Machine-readable source lineage is in source-registry.json.

Decision boundary

ThinkingCap is the selected resident Heavy quality challenger because it has the highest confirmed stable score in the current protocol-v2 sample. Nemotron 3 Super remains the better matched-1K-budget and latency result. The evidence does not establish a universal model ranking, a production promotion, five simultaneous 262K sessions, or the advertised one-million-token extension.