ToPo-ToPo/diffusiongemma-26B-A4B-it-mlx-8bit

MLX 8bit conversion of google/diffusiongemma-26B-A4B-it (mlx-vlm). Block-diffusion LM built on Gemma 4 (25.2B total / 3.8B active, MoE 128+1 experts, vision).

Provenance (self-converted)

• Source: google/diffusiongemma-26B-A4B-it (license: apache-2.0)
• Tool: mlx-vlm 0.6.9 mlx_vlm.convert (8bit affine, group_size=64), ~8.674 bpw
• model_type: diffusion_gemma is supported natively by mlx-vlm 0.6.9; no patch needed.
• chat_template.jinja is not the base repo's copy: it is the patched Gemma 4 Canonical Chat Template from ToPo-ToPo/gemma-4-26B-A4B-it-mlx-4bit, which suppresses the thinking channel when enable_thinking is false (otherwise the literal word thought leaks into the answer). Thinking is off by default. Weights are unaffected — restore the base repo's template for stock behaviour.

Usage

from mlx_vlm import load
model, processor = load("ToPo-ToPo/diffusiongemma-26B-A4B-it-mlx-8bit")

Diffusion generation takes its own flags:

python -m mlx_vlm generate --model ToPo-ToPo/diffusiongemma-26B-A4B-it-mlx-8bit \
  --prompt "Why is the sky blue?" \
  --max-tokens 256 --max-denoising-steps 48 --diffusion-sampler entropy-bound