AX-Qwen3.5-0.8B-decision-MLX-AXQ-MXFP8-MTP — 9.17 BPW measured main

An AXQuant (AXQ) mixed-precision MLX checkpoint for Apple Silicon, converted directly from the BF16 source model. The language path is quantized while the multi-token-prediction (MTP) head and vision tower are preserved at BF16 in the checkpoint (or a bound sidecar when present).

Development evidence — not a certified AXQuant release. This package has conversion and artifact-integrity records, but it does not publish measured quality, long-context, kernel-speed, or MTP-speed evidence. Do not interpret the AXQ product label as a benchmark claim.

Model details

Property	Value
Base model	mlx-community/Qwen3.5-0.8B-decision
Source revision	36e166ca10c4d16471a231d66cfc22d80039fe34
Product family	qwen3.5
Source architecture	Qwen3_5ForConditionalGeneration (dense); text path optimized
Main-model parameters	852.99M logical parameters
Quantizer	AXQuant 2.1.0
Hub budget class	MXFP8
AXQuant base precision class	16p0bpw
Planned storage-adjusted BPW	9.9741
Measured main-model BPW	9.1693
Measured total BPW, including MTP	9.3292
Safetensors weight size	1.02 GB
Approximate complete download	1.05 GB
Configured maximum context	262,144 tokens; practical limits depend on unified memory
Primary MLX runtime	MLX-LM
AX Engine native execution	Not established; no validated native manifest is included
MTP present	True
Vision present	True
Audio present	False

This repository contains MLX Safetensors. It does not contain PyTorch or GGUF weights.

Choosing an AXQ pack

AXQ names describe a storage-budget product class, not one uniform precision applied to every tensor. Protected tensors remain at higher precision, so the exact measured BPW is authoritative. In particular, a 6bit-named mixed plan may retain 4bit as its base precision while selecting 6-bit, 8-bit, or BF16 for other tensors to meet an approximately 6-BPW total budget. Protection floors can also raise a 4bit-named pack close to (or above) a 6bit budget on small or heavily protected models. When that collapse happens, AutomatosX does not publish a separate misleading 4bit sibling for that base.

Sibling	Intended trade-off
MXFP4 sibling	Same AXQ role policy at the other MLX block format. Vision, MTP, and norms stay BF16. The LoRA was merged first.

See the AutomatosX collections for the family catalog, or the complete index.

Download

python -m pip install -U huggingface_hub
hf download AutomatosX/AX-Qwen3.5-0.8B-decision-MLX-AXQ-MXFP8-MTP --local-dir ./AX-Qwen3.5-0.8B-decision-MLX-AXQ-MXFP8-MTP

Allow at least 1.05 GB of free disk space. Pin the resulting Hub commit in reproducible deployments rather than relying indefinitely on main.

Decision model

This checkpoint scores a joint schema head. It is not a chat model. mlx_lm.generate loads the backbone and produces text that is not the model's output. The published rank-16 LoRA (scale 1.0) was merged into the BF16 language weights before quantization. The adapter file is not in this pack.

joint_head.safetensors is the original BF16 joint schema head, copied unchanged, with joint_head_config.json and decision_config.json. Language attention and MLP use this pack's AXQ format. Token embeddings are 8-bit affine. Vision, MTP, and norms stay BF16. MTP acceleration was not measured. This development pack does not repeat the upstream OptiQ decision scores. The artifact records MLX 0.32.1 and MLX-LM 0.31.3 from conversion.

AX Engine status

This package does not include a validated native model-manifest.json, so AX Engine execution is not established by this release. The AX Engine fields in axquant_runtime.json describe the intended compatibility contract, not observed runtime evidence. Use the architecture-specific MLX runtime path above. The artifact records AX Engine version 7.6.3, but version discovery alone is not a runtime check.

Use the packaged Qwen MTP head with oMLX or MTPLX

This repository keeps the Qwen MTP head in mtp.safetensors; stock MLX-LM does not load that sidecar by itself. Download the complete repository to a writable local directory before using a sidecar-aware runtime.

In oMLX 0.6.3rc2 or newer, add the local directory, open Model Settings, choose Import MTP side-car, and then enable Lightning MTP. The import changes only the local copy so the MTP tensors become visible through the checkpoint index.

MTPLX can consume the packaged sidecar directly:

mtplx quickstart \
  --model ./AX-Qwen3.5-0.8B-decision-MLX-AXQ-MXFP8-MTP \
  --profile stable \
  --depth 1 \
  --reasoning off

mtplx_runtime.json declares the canonical qwen3-next-mtp execution contract. This enables strict runtime discovery; it does not extend AXQuant quality, exactness, or speed certification to oMLX or MTPLX.

Quantization layout

Main-weight precision	Parameters	Share
8bit	751.89M	86.08%
bf16	121.54M	13.92%

• Quantization methods: affine, bf16.
• Group sizes used by quantized assignments: 32.
• MTP sidecar: 15 tensors, 20.45M parameters, 0.04 GB, BF16.
• Vision sidecar: 153 tensors, 100.59M parameters, 0.20 GB, BF16.
• Vision weights: protected BF16 sidecar.
• Optimization scope: text-path.
• Support tier: convertible.

BF16 sidecars, when present, are included in total download size. Their presence does not by itself establish MTP acceleration or vision-language quality.

Evidence and validation status

Check	Status
Planning evidence	architecture_prior
Calibration	none; the allocation is based on architecture priors
Quantizer execution	187/187 recorded module conversions succeeded; 0 fallbacks
AX Engine native manifest	not included
Quality versus BF16 or uniform baselines	Not published; no quality-retention claim
MTP acceptance and speed	not measured; no MTP speedup claim
AX Engine kernel evidence	unmeasured
Vision-language quality	Not evaluated or claimed; vision tensors are preserved at BF16
Speech-recognition quality	Not applicable
Long-context quality	262,144-token capacity is config metadata, not a validated claim
Release certification	Not certified; formal AXQuant M0-M8 gates are not closed

Intended use and limitations

• Intended for local development and evaluation on Apple Silicon with MLX-compatible runtimes.

• No minimum unified-memory figure is claimed; loadability depends on model size, context length, KV-cache policy, runtime buffers, and other processes using unified memory.

• Architecture-prior allocation is not measured sensitivity. It must not be presented as measured model quality.

• MTP requires a sidecar-aware runtime. oMLX/MTPLX discovery compatibility does not establish exactness or speed certification for those runtimes.

• Vision weights are preserved at BF16, but this release does not claim validated VLM quality.

• The configured context window can require substantially more memory as the KV cache grows.

• AX Engine execution is not established because this package has no validated native manifest.

• Upstream capabilities, limitations, biases, and responsible-use guidance still apply.

Provenance and audit files

• axquant_manifest.json: package identity, byte accounting, runtime contract, software versions, and file checksums.
• axquant_plan.json: per-tensor precision decisions and planning evidence.
• axquant_quantizer_execution.json: conversion coverage and fallback records.
• axquant_runtime.json: declared AX Engine and MLX compatibility metadata; runtime checks remain separate evidence.
• axquant_mtp_sidecar_manifest.json: MTP tensor provenance.
• axquant_vision_sidecar_manifest.json: protected vision tensor provenance.

All published provenance uses repository-relative paths. Local source paths are stripped before publication. The checkpoint was converted from BF16 rather than re-quantized from an OptiQ artifact. If an OptiQ repository is published separately, it uses a different quantizer and should not be assumed to have identical BPW or quality.

License

The checkpoint follows the upstream model license where applicable (often Apache License 2.0). See LICENSE and the mlx-community/Qwen3.5-0.8B-decision model card for license terms, model limitations, and responsible-use guidance.