Hacker News front pageAug 20, 2026, 21:24 UTC+8

DiffusionGemma Technical Report: High-Speed Text Generation via Discrete Diffusion

DiffusionGemma Technical Report

Google's DiffusionGemma is an experimental open-weight LM that generates text by iteratively refining 256-token blocks in parallel, hitting roughly 1,500 tokens/sec on a single H100. It is fine-tuned from the MoE Gemma 4 (3.8B active, 25.2B total) using under 10% of the original training token budget. A two-stage pipeline—supervised fine-tuning for bidirectional denoising, then RL with sampler distillation—improves both quality and speed. The model sets a new Pareto frontier for speed vs. capability, retains thinking mode, multimodal inputs, and long-context support, and can still do autoregressive decoding with minor degradation.

Why it matters: DiffusionGemma applies diffusion to text generation with 256-token parallel blocks, hitting ~1,500 tok/s on a single H100 — faster than speculative decoding. Not pushing past 84 because it's still a tech report with no product path or real-world deployment data yet.

Read the original ↗Export Markdown

Take

This one's worth a click because it moves diffusion-based text generation from a research curiosity to an open-weight model you can actually run. DiffusionGemma isn't trained from scratch—it's fine-tuned from the MoE Gemma 4 (3.8B active, 25.2B total params) using under 10% of the original training token budget. The two-stage pipeline does supervised fine-tuning for bidirectional denoising, then RL with sampler distillation to improve both quality and inference speed.

The headline number: roughly 20 tokens per forward pass, hitting ~1,500 tok/s on a single H100. That's faster than autoregressive models even with state-of-the-art speculative decoding. It also keeps thinking mode, multimodal inputs, and long-context support, and can fall back to autoregressive decoding with minor quality loss.

Where I'd hold back: the paper says results are averaged across their full eval suite, but doesn't break out specific benchmarks for long-form generation or complex reasoning. The speed is real, but if quality tanks on certain tasks, this is more of a high-throughput batch generation tool than a general-purpose chatbot. Also, diffusion sampling steps matter a lot for quality—how few steps you can get away with in practice is still an open question until people start running it.

Source: Hacker News front page · arxiv.org
Published: Aug 20, 2026, 21:24 UTC+8 · Aug 20, 2026