Qwen3.8-27B-OptiQ-4.5bpw-mixed

Model-card update:2026-09-14.

Status

Historical published candidate, not a current B/C pick. Its recorded recipe and tests are retained below; it has not received the current two-pick C84 runtime screen or a new predictor certification.

Conversion and attribution

MLX OptiQ KL-sensitivity 4/8-bit conversion of unsloth/Qwen3.8-27B, with the vision tower retained. Historical serving metadata reports 27B VLM parameters, 4.98 effective bits/weight, 18.44 GB weights and 152 eight-bit/346 four-bit quantized modules. The 4.5bpw repository name is not the measured effective precision. The intermediate conversion print 5.485bpw/17.6 GB belongs to a different packaging stage and must not replace the final manifest accounting.

Packed low-bit tensor counts can understate true parameter counts in hub badges. Weight-file sizes, effective bits/weight and total MLX runtime peak are different measurements.

Historical results and limits

The 2026-08-17/19 predictor-OFF t0.6 code screen recorded HumanEvalPlus 42/48 graded, 42/50 strict and MBPPPlus 40/50. Two HumanEvalPlus timeouts came before the repaired C28 transport/orphan-cascade handling; these are historical, compromised shallow-screen data, not evidence of precision preventing runaways. No post-fix shallow re-run of this exact recipe is claimed.

M25,2026-08-29, opencode 22-item sets at t0.6/template xhigh, predictor OFF:18/22 Python and 16/22 Go; comparison against Qwen3.8-27B-mlx-uniform-4bit at20/22 and16/22 was a historical WASH, not promotion.

The old capacity ladder recorded 33.72 GB MLX peak, 1977 s prefill and 10.53tok/s decode at nominal 256K, retrieval co-score 1.0. It used an old runtime/predictor-OFF protocol and historical 46 GB flag. Rough 48 GB is now guidance, not a strict rejection cutoff; old capacity/speed must not describe today's two-pick runtime. No general causal claim that architecture alone determines decode speed is retained.

T0.6 and omitted reasoning_effort retain the historical template default (xhigh in those runs); no medium-effort or MTP certification exists for this recipe. The bundled original MTP file is inert under normal target loading; do not enable it by copying a sibling's recommendation. Vision tower presence is recorded, without a new quality gate.

Predictor

Predictor OFF: no external MTP recommendation is configured or certified for this exact recipe. A bundled head or MTP tag alone is not evidence to enable speculation.

Recorded historical-candidate configuration, not a new recommendation

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-OptiQ-4.5bpw-mixed
  type: vision
  hf_path: caslca/Qwen3.8-27B-OptiQ-4.5bpw-mixed
  max_kv_cache_size: 262144
  kv_prealloc_tokens: 262144
  kv_quant_scheme: turboquant
  kv_bits: 4
  quantized_kv_start: 0
  prefill_step_size: 512
  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
  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-OptiQ-4.5bpw-mixed. 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 14 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.