DiffusionGemma: The Developer Guide

Summary

DiffusionGemma is an experimental text-generation model and developer guide covering serving, customization, and non-autoregressive generation. Built on the Gemma 4 backbone, its 26B Mixture of Experts architecture activates 3.8B parameters during inference, refines 256-token canvases in parallel, and shifts the bottleneck from memory bandwidth to compute. Block autoregressive diffusion commits denoised 256-token blocks to the KV cache, combining bidirectional denoising with causal prefill for longer sequences; vLLM integration provides an OpenAI-compatible local server. A Sudoku fine-tuning example using Hackable Diffusion reports improvement from roughly 0% base-model success to 80%, while SFT reduces the demonstrated solve from 48 to 12 steps through adaptive stopping. The guide also identifies bidirectional attention and re-noising as mechanisms for global constraint handling and self-correction.

Takeaways

1. 1

   DiffusionGemma's 26B MoE architecture activates 3.8B parameters during inference, enabling quantized deployment within 18 GB of VRAM.

2. 2

   The model combines bidirectional attention during denoising with causal incremental prefill, allowing canvas-wide context, re-noising for self-correction, and sequential handling of longer outputs.

3. 3

   On Sudoku, the base model reports approximately 0% success, while a simple JAX SFT recipe raises correctness to 80%; the demonstrated fine-tuned run finishes in 12 steps versus 48 for the base model.