---
license: apache-2.0
license_link: https://ai.google.dev/gemma/apache_2
base_model: google/diffusiongemma-26B-A4B-it
pipeline_tag: image-text-to-text
library_name: mlx
tags:
  - mlx
  - apple-silicon
  - diffusion-gemma
  - mixture-of-experts
  - quantized
  - 4-bit
  - mirror
  - automatosx
  - ax-engine
---

# AX DiffusionGemma 26B-A4B IT MLX 4-bit

> **Parameter count:** approximately 25.82B logical parameters (26B total,
> approximately 4B active per token). `4-bit` is the quantization precision,
> not a 4B model-size claim.

> **Transparent upstream mirror — AutomatosX made no model changes.**

This is a revision-pinned, transparent mirror of
[mlx-community/diffusiongemma-26B-A4B-it-4bit](https://huggingface.co/mlx-community/diffusiongemma-26B-A4B-it-4bit)
at commit `a7a81407613811e8ba63af92ac0d852b809e191f`.

AutomatosX did not train, fine-tune, merge, convert, quantize, re-quantize, or
otherwise alter the weights, configuration, tokenizer, processor, chat
template, or generation settings. The MLX conversion and quantization were
performed upstream by MLX Community with `mlx-vlm` 0.6.3. AutomatosX adds only
this mirror documentation, the preserved upstream model card, a copy of the
declared license, machine-readable provenance, and a `total_parameters` field
in the shard index so the Hub does not mistake packed 4-bit storage for the
logical model size.

## Model details

- Base model: [google/diffusiongemma-26B-A4B-it](https://huggingface.co/google/diffusiongemma-26B-A4B-it)
- Upstream MLX source: [mlx-community/diffusiongemma-26B-A4B-it-4bit](https://huggingface.co/mlx-community/diffusiongemma-26B-A4B-it-4bit)
- Format: MLX Safetensors for Apple Silicon
- Architecture: `DiffusionGemmaForBlockDiffusion`, mixture of experts
- Quantization: upstream 4-bit affine, group size 64, with the upstream 8-bit exceptions unchanged
- Weight shards: 4, totaling 16,543,055,405 bytes
- Generation: block-autoregressive discrete diffusion with a 256-token canvas
- MTP: not applicable; this is not an autoregressive next-token model

## Download

~~~bash
hf download AutomatosX/AX-DiffusionGemma-26B-A4B-IT-MLX-4bit \
  --local-dir ./AX-DiffusionGemma-26B-A4B-IT-MLX-4bit
~~~

## Use with MLX-VLM

~~~bash
pip install -U mlx-vlm
python -m mlx_vlm.generate \
  --model AutomatosX/AX-DiffusionGemma-26B-A4B-IT-MLX-4bit \
  --max-tokens 100 \
  --temperature 0.0 \
  --prompt "Describe this image." \
  --image <path_to_image>
~~~

## Try with AX Engine

[AX Engine](https://github.com/defai-digital/ax-engine) has an experimental
native MLX path for DiffusionGemma. Its download workflow generates the local
runtime manifest without changing the published model artifacts:

~~~bash
ax-engine serve AutomatosX/AX-DiffusionGemma-26B-A4B-IT-MLX-4bit \
  --download --port 31418
~~~

DiffusionGemma support in AX Engine is experimental and is not recommended for
production use. Its first visible output is a committed diffusion block, so
ordinary autoregressive TTFT and decode-throughput comparisons do not apply.

## Mirror policy and provenance

`UPSTREAM_README.md` preserves the original MLX Community model card.
`ax_provenance.json` pins the source commit and records SHA-256 values and sizes
for every mirrored artifact. All weight shards and all source artifacts other
than the replaced model card and shard index are byte-identical to the pinned
upstream revision. In the index, `weight_map` and `total_size` remain unchanged;
only the logical `total_parameters` metadata field was added.

This repository exists for revision-pinned cataloging and convenient discovery;
it does not represent a new model, conversion, quantization, or benchmark claim
from AutomatosX.

## License and responsible use

The upstream MLX repository and Google base model declare Apache License 2.0.
See `LICENSE`, the preserved upstream card, and the
[Google base-model card](https://huggingface.co/google/diffusiongemma-26B-A4B-it)
for usage limitations, safety guidance, and responsible-use considerations.