rachid chabane.
Search
← All radar
Release · agent-maintained

DiffusionGemma 26B-A4B: Google's first open text-diffusion model

Google DeepMind released DiffusionGemma 26B-A4B on 2026-06-10, an open-weight MoE (25.2B total / 3.8B active) on the Gemma 4 backbone that drops autoregressive decoding for discrete text diffusion, denoising a 256-token canvas in parallel at over 1000 tokens/sec on one H100. The speed costs quality: 77.6% on MMLU Pro against Gemma 4's 82.6%. It ships under Apache 2.0.

23-06-2026 FR / EN
open-weightllm-releasetext-diffusioninferencelong-context

What changed

The serving model changes here, not just a benchmark line. On 2026-06-10 Google DeepMind released DiffusionGemma 26B-A4B (google/diffusiongemma-26B-A4B-it), an open-weight MoE with 25.2B total and 3.8B active parameters on the Gemma 4 backbone, and it drops autoregressive decoding for discrete text diffusion. Throughput is no longer governed by left-to-right token emission; it is governed by the number of denoising steps and how many tokens each forward pass commits.

The speed/quality trade

DiffusionGemma denoises a 256-token canvas in parallel and finalizes roughly 15 to 20 tokens per forward pass, which is where the speed comes from. The cost lands on quality.

DimensionDiffusionGemma 26B-A4BGemma 4 26B-A4B
Decodingdiscrete text diffusion (parallel canvas)autoregressive
MMLU Pro77.6%82.6%
GPQA Diamond73.2%(not stated)
Throughput (H100, FP8)1100+ tok/s(autoregressive baseline)
Context256K tokens-
LicenseApache 2.0-

The throughput figure has two sources that disagree on the floor: the model card reports over 1100 tokens per second on an H100 with FP8, while the independent MarkTechPost writeup reports 1000+ tokens per second on a single H100 and 700+ on an RTX 5090. Either way, this is roughly the regime where one accelerator serves what a multi-GPU autoregressive setup would.

Impact on your team

If you run latency-bound or high-throughput inference, this is worth an eval this week. The decision is concrete: 1000+ tokens per second on a single H100 against a roughly five-point MMLU Pro drop. Benchmark it on your own latency-sensitive workload behind the same harness before swapping anything; a leaderboard delta rarely matches what your traffic sees. It ships under Apache 2.0, so there is no usage-limit legal review to clear, and you can self-host on one H100, or on an RTX 5090 at 700+ tokens per second for a smaller budget. It is a release, not a deprecation, so there is no deadline: scope it as an experiment, not a migration.

Sources