DiffusionGemma — experimental open diffusion language model from Google Gemma

TL;DR

Google Gemma’s official handle announces DiffusionGemma, framed as an experimental open Gemma-family model released under Apache 2.0 that explores diffusion-style text generation: instead of producing tokens one at a time autoregressively, it denoises entire blocks of text in parallel. The post is a launch teaser only (video attached, “what’s new with DiffusionGemma 👇” thread) — no benchmarks, no model card, no parameter count in the tweet body at filing time. This is Google’s first public Gemma-line entry into the diffusion-language-model thread that has, until now, been dominated by Apple (DiffuCoder), NVIDIA (Nemotron-Labs-Diffusion), Inception Labs (Mercury), Tencent (WeDLM), and Zyphra (ZAYA1-Diffusion-Preview).

Key claims

• DiffusionGemma is released as an open model under Apache 2.0 [tweet body].
• The model generates entire blocks of text simultaneously rather than sequential token-by-token decoding [tweet body].
• Position is “experimental” — Google Gemma’s framing, not a flagship release [tweet body].

Method

The tweet itself does not describe the architecture, training data, or sampling procedure — only that the approach moves “beyond sequential, token-by-token processes to generate entire blocks of text simultaneously,” consistent with how this page already describes masked / absorbing-state discrete diffusion language models (see TL;DR of Diffusion Language Models). An attached video and an in-thread reveal (“Here’s what’s new with DiffusionGemma 👇”) presumably contain method detail; the thread is not retrievable from the tweet view alone at filing time. Whether this is a from-scratch dLLM or an AR-base continued-pretrain from a Gemma checkpoint (the dominant recipe per Diffusion Language Models) is unspecified.

Results

None reported in the tweet. Filing a placeholder pointer; a follow-up tech report / model card from Google should be filed as a separate artifact when available, with this page kept as the tweet-as-pointer.

Why it’s interesting

First diffusion-LM entry from the Gemma program — joins Introspective Diffusion Language Models (Apple’s I-DLM converted from Qwen3-8B), Nemotron-Labs-Diffusion: A Tri-Mode Language Model Unifying Autoregressive, Diffusion, and Self-Speculation Decoding (NVIDIA’s tri-mode AR+diffusion+self-speculation), and ZAYA1-8B-Diffusion-Preview: Converting an AR MoE to Diffusion on AMD (Zyphra’s AR-MoE→diffusion conversion) in the “major lab ships a dLLM” thread. The Apache-2.0 license puts it directly into the open-weights cohort tracked by Open foundation-model releases alongside Gemma 4 (Gemma 4 12B — encoder-free unified multimodal model with 35M vision embedding (Google Gemma)), strengthening the read that Google’s Gemma program is moving past flagship-only releases to ship architectural experiments openly. Counterbalances the framing in the Diffusion Language Models open-questions section that “frontier-quality dLLMs have so far come from labs without flagship AR LLMs” — Google now has both.

See also

• Diffusion Language Models — DiffusionGemma is a new datapoint on the dLLM thread; pending tech detail to know which sub-axis (throughput, serving, RL, scaling) it sits on.
• Open foundation-model releases — Apache-2.0 open release fits the Gemma program’s open-weights pattern.
• Gemma 4 12B — encoder-free unified multimodal model with 35M vision embedding (Google Gemma) — Companion Gemma announcement; Gemma 4 12B was unified-multimodal, DiffusionGemma is the architectural-experiment lane.
• Nemotron-Labs-Diffusion: A Tri-Mode Language Model Unifying Autoregressive, Diffusion, and Self-Speculation Decoding — NVIDIA’s tri-mode dLLM; closest filed comparable as a major-lab experimental dLLM release.
• Introspective Diffusion Language Models — Apple’s I-DLM; closest filed comparable for AR→diffusion conversion at the 8B scale, which DiffusionGemma may or may not follow.