AX-Qwen3.6-35B-A3B-MLX-AXQ-MXFP4-MTP

AXQuant MXFP4 development conversion of the official Qwen3.6-35B-A3B BF16 checkpoint. This is a development artifact, with no Tier 1, Tier 2, quality, MTP acceptance, or speed certificate.

Source and storage

  • Source: Qwen/Qwen3.6-35B-A3B.
  • Immutable source revision: 995ad96eacd98c81ed38be0c5b274b04031597b0.
  • AXQuant version: 2.3.0.
  • Main-model storage: 4.6460 BPW, measured from serialized files.
  • Total storage including MTP: 4.9127 BPW.
  • Safetensors download: 22.078 GB.
  • Eligible language attention, expert, and MLP projections use native MLX MXFP4, group32.
  • Vision, MTP, routers, norms, and the output head retain BF16 precision.
  • Embeddings use 8-bit affine group32 packing.
  • Manual planning is unmeasured architecture-prior evidence. MXFP8 is explicit unmeasured physical repacking, not a measured sensitivity method.

BPW includes protection floors and metadata overhead; the format name is not a claim that every tensor has the same precision. Source weights were not trained or augmented, and no competitor quantization data was used.

Runtime evidence

The exact payload passed bounded MLX-LM text generation on the factory Mac Studio M2 Ultra. Synthetic image generation was attempted but the standard MLX-VLM loader did not load the protected, unindexed vision sidecar. See runtime_smoke.json for the dependency versions, answers and payload digests.

Test environment: MLX 0.32.1, MLX-LM 0.31.3, MLX-VLM 0.6.17. Smoke results establish only the tested load/generation cases. Vision generation is not supported by this export through the default MLX-VLM loader. The BF16 vision payload is retained separately for a future explicit runtime export. Broad text, image, video, tool-use, long-context, and task-quality evaluations have not been performed.

mtp.safetensors contains the protected native MTP payload. Its presence does not establish MTP execution, acceptance, exactness, or acceleration. MTP on/off execution has not been verified for this artifact. The ordinary MLX-LM example below uses autoregressive text generation.

axquant_compatibility.json is manifest-bound static evidence. Its unsupported oMLX/MTPLX profiles remain unsupported; this publication does not promote runtime compatibility or claim native AX Engine qualification.

Text generation

Install MLX-LM on Apple Silicon, then load the complete repository:

from mlx_lm import load, generate

model, tokenizer = load("AutomatosX/AX-Qwen3.6-35B-A3B-MLX-AXQ-MXFP4-MTP")
prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "What is 6 times 7?"}],
    tokenize=False, add_generation_prompt=True, enable_thinking=False,
)
print(generate(model, tokenizer, prompt=prompt, max_tokens=64))

Pin the immutable Hub revision for reproducible deployments. Actual memory use depends on model loading, context, vision input, and runtime settings.

Provenance and license

axquant_manifest.json binds output files, source identity, the sanitized plan, quantizer execution, protected sidecars, and static compatibility. The original Apache-2.0 license is included in LICENSE; NOTICE identifies the quantization changes. Existing historical certificates do not bind this new export and are not reused.

Selected for the AX BI collection and AX Work collection.

Downloads last month
215
Safetensors
Model size
35B params
Tensor type
BF16
·
U32
·
MLX
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for AutomatosX/AX-Qwen3.6-35B-A3B-MLX-AXQ-MXFP4-MTP

Quantized
(1,015)
this model

Collections including AutomatosX/AX-Qwen3.6-35B-A3B-MLX-AXQ-MXFP4-MTP