Qwen3.6-27B-Opus-Distill-OptiQ-4bit-mtp-drafter

Qwen3.6-27B-Opus-Distill-OptiQ-4bit-mtp-drafter

Model-card update:2026-10-30.

This repository is an external MTP predictor for Qwen3.6-27B-Opus-Distill-OptiQ-4bit, not a standalone chat model. Fourth choice on the operator-approved coding menu (B4); not a current research/design pick. Temperature 0.3 with the certified native MTP companion remains configured. Ordinary correctness can conceal expensive nonconvergence on math/research tasks.

Predictor-ON certification on the current runtime, 2026-10-30

MTP ON is the CERTIFIED default for this target on the current runtime (MLX-VLM 1c7c8fa1, MLX-Serve 1ddf6f84, MLX/Metal 0.32.4), measured on an M5 Max 128 GB machine on AC power. Earlier sections describe earlier runtimes and stay as recorded.

• Exactness: a full speculative-decoding exactness campaign — forced-token ON vs OFF at short, 1K, 8K, 32K and 64K context (maximum total-variation distance 0), greedy identity on 24 math / code / judge items, cross-load identity, and multi-turn continuity — PASS (wall 2.8 h).
• Normal-length decode (math500, 20 seeded-random items, deployed sampling, thinking on, paired by item): outputs byte-identical ON vs OFF on 20 / 20 items; decode ON / OFF 1.74×[1.71,1.77] (geometric mean of per-item ratios, 95% item bootstrap), token-weighted 38.9 vs 22.7 tok/s, mean 14.2K (max 82.5K) generated tokens; draft acceptance 0.89. One pair; descriptive.
• At the 262144-token cap (a 260096-token prompt, OFF / ON / ON / OFF, ≥ 10 minutes idle between arms): prefill-spike peak 28.14 → 28.46 GB, decode peak 27.95 → 29.60 GB; prefill unchanged; decode 10.73 → 10.09 tok/s (−6%; directional — the two OFF replicates spread 11.19 / 10.27) with MTP ON. Near a full context the verification step's extra full-context attention outweighs the accepted drafts; the crossover lies between about 82K and 262K tokens and was not located.

Limits: speed figures come from one machine and one pair per state at normal length; math500 accuracy is equal by construction (identical outputs), not a quality ranking.

Predictor lineage and identity

One-layer native qwen3_5_mtp extraction of the matching target’s MTP head, affine int4/group 64, roughly 300 MB plus tokenizer. Historical serving used two drafted tokens per round. Pair only with the matching target artifact; acceptance and quality do not transfer to another fine-tune or quantization.

Historical predictor and quality limits

M6b, 2026-08-25, HumanEvalPlus 63 paired tasks, temperature 0.3/TQ4: ordinary 95.24% ON versus 93.65% OFF, delta+1.6 pp[0,+4.8], historical±5 pp equivalence; acceptance 0.923 and approximately 2× decode. This is the recorded predictor certification, not a claim that every later axis is equivalent.

C63 matched Math500,2026-09-11, predictor OFF: ordinary 99%, strict 93%, 94% convergence at 81920 budget, n 100; five repetition runaways plus one budget hit consumed 37% of 21.72 h. The strict score preserves the cost of nonconvergence despite correct final answers. This is why ordinary accuracy alone must not advertise this model as an unrestricted research default.

M39,2026-09-12, predictor OFF:18/20 on a 20-photo COCO description/self-check gate; one person-description mistake and one missed face-wiping detail. Three descriptions were manually inspected. Limited visual functionality is established, not comprehensive image understanding. No current C84 runtime screen for this model is claimed.

Recommended current setting

Save this as models.yaml in an environment with the compatible fork revisions described below. The IDs name the target and, where configured, its exact companion; weight quantization and attention-KV precision are separate settings.

manager_port: 8000
mlx_port: 8091
startup_timeout_seconds: 300
models:
- name: Qwen3.6-27B-Opus-Distill-OptiQ-4bit
  type: vision
  hf_path: caslca/Qwen3.6-27B-Opus-Distill-OptiQ-4bit
  max_kv_cache_size: 262144
  kv_prealloc_tokens: 262144
  kv_quant_scheme: turboquant
  kv_bits: 4
  prefill_step_size: 512
  quantized_kv_start: 0
  generation_defaults:
    temperature: 0.3
    top_p: 0.95
    top_k: 20
    min_p: 0.0
    presence_penalty: 0.0
    max_tokens: 102400
    thinking_budget: 81920
    enable_thinking: true
  draft_kind: mtp
  draft_model: caslca/Qwen3.6-27B-Opus-Distill-OptiQ-4bit-mtp-drafter
  on_demand: true

Start from the compatible installed fork environment without additional worker overrides. This command clears cache/predictor environment overrides, disables APC and selects two retained sessions with one resident model:

env -u APC_ENABLED -u KV_BITS -u KV_QUANT_SCHEME -u KV_GROUP_SIZE \
  -u KV_KEY_BITS -u KV_VALUE_BITS -u KV_KEY_SCHEME -u KV_VALUE_SCHEME \
  -u QUANTIZED_KV_START -u MLX_VLM_SESSION_SHRINK_ON_RETIRE \
  -u MLX_VLM_DRAFT_KIND -u MLX_VLM_DRAFT_MODEL -u MLX_VLM_DRAFT_BLOCK_SIZE \
  MLX_VLM_CACHE_SESSION_MAX=2 MLX_SERVE_CONFIG=models.yaml mlx-serve start

The model loads on demand. An OpenAI-compatible client sends requests to http://localhost:8000/v1 with model ID Qwen3.6-27B-Opus-Distill-OptiQ-4bit. Omitted sampling fields inherit generation_defaults; thinking is enabled. The output cap is 102400 and thinking budget 81920. Near the context limit the server resolves a smaller budget from remaining space; a configured 262144 cap does not imply 262144 input tokens plus another 102400 output tokens.

The full active KV floor equals the context cap; keep prefill 512. kv_bits: 0 means unquantized native16 attention KV, not unquantized model weights or a newly measured IEEE-FP16 dtype. TurboQuant is inactive when bits are 0 even if its scheme name remains configured. Idle retirement is explicitly true only on Qwen3.8-27B-Fable-Distill-OptiQ-4.5bpw-mixed; omitted cache_session_shrink preserves the worker default OFF only when MLX_VLM_SESSION_SHRINK_ON_RETIRE is absent, as in the command above. Suffix decoding stays OFF. These structural settings target the 64 GB evaluation machine, not arbitrary hardware.

Runtime and evidence scope

The final C84 environment uses MLX-VLM 522671c4bebc5ff492d465a1d0e6a14251f18260, MLX-Serve b632280709f771972bffbaf3231e996e8a89f4e8, and MLX/Metal 0.32.2. Those fork commits are published. Live validation of this exact final bundle covers the two current picks only; other cards retain historical model evidence and source-level compatible configuration, not a newly run model smoke. Older experiments used their recorded earlier runtime and cache mode.

Dates, original source revisions, per-axis observations and evidence-file hashes are recorded in the accompanying evaluation evidence alongside this card. The tables here are self-contained: they do not rely on unpublished stack commits or links to future GitHub main content. Speculative decoding can change BF16 outputs; favorable task-specific intervals do not imply byte-lossless inference. Historical and current measurements are not pooled across cache modes, predictor states, budgets, corpora or runtimes.

Artifact integrity audit, 2026-09-14

All 4 audited model and support files match the corresponding local artifact by full-byte cryptographic hashing, including weights and applicable configuration/tokenizer/vision files. README, license and repository-attribute files are tracked separately from inference artifacts. This publication changes the card and its dated evidence only; model artifacts remain unchanged. Per-file hashes, audited revision and scope.