Qwen3.8-27B-mlx-uniform-4bit-mtp-drafter

Model-card update:2026-10-30.

This repository is an external MTP predictor for Qwen3.8-27B-mlx-uniform-4bit, not a standalone chat model. Second operator-approved choice for agentic coding and research/design assistance (B2/C2), at temperature 0.6, medium reasoning effort and native MTP ON. The current runtime retains TurboQuant 4 KV. The M40 MBPPPlus predictor comparison remains INCONCLUSIVE; deployment approval does not erase that limit.

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 1.6 h; the 8K straddle cell valid at prompt offset 65).
• 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 1.72×[1.66,1.79] (geometric mean of per-item ratios, 95% item bootstrap), token-weighted 47.3 vs 28.1 tok/s, mean 1.6K generated tokens; draft acceptance 0.81. One pair; descriptive.
• At the 262144-token cap (a 260096-token prompt, OFF / ON / ON / OFF, ≥ 10 minutes idle between arms): prefill-spike peak 25.02 → 25.33 GB, decode peak 25.18 → 26.06 GB; prefill unchanged; decode 11.92 → 10.38 tok/s (−13%) 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. The historical mlx-community twin relationship applies only to the explicitly compared target revisions.

Earlier predictor certification, 2026-08-30 (M6d)

At xhigh reasoning effort/TQ4,164 paired HumanEvalPlus tasks: ordinary 94.5% both states, delta 0 pp[−3,+3]; strict 93.9% ON versus 92.7% OFF. Mean acceptance 0.674 and paired mean decode 1.60×. The later medium-speed screen was an instrument check; it is now followed by the separate M40 medium-quality evidence below, rather than an enduring “not measured” claim.

M40 predictor-ON certification, 2026-09-13 — historical TQ4 configuration

These ON/OFF comparisons used the same deployed tune, matched seeds, cap 262144 and resolved thinking budget 81920. Both predictor states remain recorded; new optimization work runs in the shipped ON state. This supersedes earlier blanket statements that every benchmark always runs predictor-OFF.

Operator-approved state is MTP ON. Per-axis limits remain: a zero-discordance empirical interval is not proof of universal equivalence; n 100 nominal MDE is 12.5 pp and n 40 is 20 pp. Vision and depth drivers verified the worker flag but did not expose per-request draft counters. Those small gates do not certify arbitrary images, OCR, complex diagrams, or full-context reasoning. The photo gate uses description followed by a ground-truth-caption self-check; it is a lenient capability screen, not a ranked vision benchmark. C74 explicitly preserves the MBPPPlus INCONCLUSIVE label; pooled coding diagnostics do not replace it.

C84 actual-runtime regression screen, 2026-09-14

M5 Max 64 GB, MLX/Metal 0.32.2. BEFORE was the initially merged runtime e3bffd9a/f8f1df4. AFTER combines later upstream changes, cache lifetime/logical-trim/eager-retirement repairs, and the per-model retirement option; retirement is enabled only for Qwen3.8-27B-Fable-Distill-OptiQ-4.5bpw-mixed. This is not an isolated retirement/kernel effect or the original pre-merge/upstream comparison.

Five frozen seeded tasks per axis, 20 requests in each state for this model. All 20 pairs have identical visible answers, reasoning and completion-token counts; all 40 responses converge. Ordinary accuracy equals strict accuracy. No exclusive solves or new nonconvergence were observed.

Intervals use 10000 paired two-stage task-bootstrap resamples, seed 84; they are nominal exploratory intervals, not multiplicity-adjusted or repeated-session uncertainty. Five items per axis have nominal MDE 56 pp; empirical quality intervals[0,0] with zero discordance do not establish±5 pp equivalence. The prose review was nonblinded, one reviewer, and found shared factual/methodological weaknesses; matching answers are not proof of correctness. This bounded screen does not rerank models or replace broad native16 depth/vision certification.

Final runtime smoke covers arithmetic, Python, JSON, native tool continuation and one fixed image, five cases/six requests. It passed for both current picks. Only Qwen3.8-27B-Fable-Distill-OptiQ-4.5bpw-mixed received the additional fresh tool-pair and three-turn prefix-reuse checks.

Total short-task time 821.4→822.9 s; per-axis changes are below 1%, not a claim of zero speed cost. An earlier integrated-runtime TQ4 capacity ladder on 2026-09-13 measured 37.7970 GB peak, 1598.33 s prefill and 6.15tok/s decode at 261449 prompt tokens (MTP ON, sourcec5a6f97b). That closes the historical capacity gap but is not a fresh final-C84 full-context comparison.

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: Qwen3.8-27B-mlx-uniform-4bit
  type: vision
  hf_path: caslca/Qwen3.8-27B-mlx-uniform-4bit
  max_kv_cache_size: 262144
  kv_prealloc_tokens: 262144
  kv_quant_scheme: turboquant
  kv_bits: 4
  quantized_kv_start: 0
  prefill_step_size: 512
  draft_kind: mtp
  draft_model: caslca/Qwen3.8-27B-mlx-uniform-4bit-mtp-drafter
  generation_defaults:
    temperature: 0.6
    top_p: 0.95
    top_k: 20
    min_p: 0.0
    presence_penalty: 0.0
    max_tokens: 102400
    thinking_budget: 81920
    enable_thinking: true
    reasoning_effort: medium
  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 Qwen3.8-27B-mlx-uniform-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.