Google's DiffusionGemma: Parallel Decoding Breaks LLM Text Generation Speed Limits
bodonoghue85 · x · 2026-08-04
Google's team released a 55-page technical report on DiffusionGemma, an experimental open-weight language model based on discrete diffusion. Unlike conventional autoregressive (AR) models that decode sequentially token-by-token, DiffusionGemma iteratively refines blocks of 256 tokens in parallel, avoiding the sequential decoding bottleneck.
Core Techniques & Training
- Fine-tuning: Instead of training from scratch, DiffusionGemma was obtained by fine-tuning the Gemma 4 MoE model (25.2B total and 3.8B activated parameters).
- Efficient Pipeline: Uses a two-stage compute-efficient pipeline requiring less than 10% of the original AR model's token budget. Stage 1 uses SFT for bidirectional denoising; Stage 2 applies RL and sampler distillation to improve generation quality.
Performance
The report highlights that diffusion can bring a very different throughput frontier. It is not limited to batch size 1 (bs=1) scenarios, demonstrating exceptional high-speed text generation capabilities.
Related event: DeepMind Releases DiffusionGemma for Fast Text Generation(5 posts)→
More from Models
- Ditching Claude Code: Migrating Complex Workflows to Open-Weight Models — TheZachMueller · 2026-08-04
- Hands-on Comparison: MMH3 Outperforms Google Omni Flash — Ok-Act-9620 · 2026-08-04
- Managing Long Contexts: Devs Share Workarounds for LLM 'Lost in the Middle' — Entire-Ship8618 · 2026-08-04
- Kimi K3 Speculative Decoding Model Hits 600K Downloads, Boosts AMD MI355X Throughput — bookwormengr · 2026-08-04
- Kimi K3 Requires Full Message History: Most Users Are Using It Wrong — NielsRogge · 2026-08-04
- DiffusionGemma Tech Report: Parallel Decoding Breaks LLM Inference Speed Limits — bodonoghue85 · 2026-08-04