DiffusionGemma 26B-A4B IT

MLQ News reports on the June 10, 2026 Google DeepMind release of DiffusionGemma, a 26B open-weights model that generates text via discrete diffusion — the same technique underpinning image generators like Stable Diffusion — producing up to 256 tokens in parallel per forward pass. The article confirms Apache 2.0 distrib The piece documents that DiffusionGemma achieves over 1,000 tokens per second on a single NVIDIA H100, roughly four to five times faster than comparable autoregressive models. Output quality is reported as lower than standard Gemma 4 on benchmarks such as MMLU and coding tests, with Google positioning the model as expe

Google released DiffusionGemma as an open model on June 10, 2026, positioning it as a diffusion-based language model built on the Gemini Diffusion research line. Unlike autoregressive models that emit tokens sequentially, DiffusionGemma iteratively refines outputs across the full sequence, a paradigm translated from im The article details concrete architecture: a Mixture-of-Experts model with roughly 25.2 billion total parameters and 3.8 billion active parameters (the "26B-A4B" configuration). In benchmark comparisons against Gemma 4 31B, Gemma 4 26B A4B, and Gemma 4 12B, DiffusionGemma reportedly produces dramatically more output to

Google AI, including Google DeepMind researchers, released DiffusionGemma on June 10, 2026 as an experimental open-weights model that uses text diffusion rather than standard autoregressive decoding for generation. The model is shipped under the Apache 2.0 license, making it available to developers and researchers expl MarkTechPost confirms DiffusionGemma's architecture as a 26B-parameter Mixture-of-Experts text-diffusion model and highlights Google's claim of up to 4x faster generation compared with conventional autoregressive approaches. The article frames the release as part of an emerging line of diffusion language models and pos

Google DeepMind released DiffusionGemma, a 26B-parameter Mixture-of-Experts open-weights model released under Apache 2.0, on June 10, 2026. The model uses discrete text diffusion to generate 256-token blocks in parallel per forward pass, shifting away from token-by-token autoregression. With 3.8B active parameters duri Google's headline performance claims include 1,000+ tokens per second on a single NVIDIA H100 and 700+ tokens per second on an RTX 5090, described as roughly four to five times faster than comparable autoregressive models. The source explicitly notes these speedups are designed for local and low-concurrency inference,

The ModelScope listing for google/diffusiongemma-26B-A4B-it carries Google DeepMind authorship and serves as the first-party model card mirror under Apache 2.0 license, with 25.82B parameters in the Transformers/Safetensors/PyTorch format. DiffusionGemma is described as a generative model built by Google DeepMind based Key technical capabilities explicitly documented include configurable Thinking Mode for reasoning, native system prompt support inherited from Gemma 4, and specific optimization for small-batch, low-latency, single-accelerator inference. The listing indicates the model is positioned to reduce the sequential bottlenecks