Google releases DiffusionGemma 26B, open-weight model generates 500+ tokens/second

Google has released DiffusionGemma 26B, an open-weight text generation model licensed under Apache 2. The model is based on Google's previously experimental Gemini Diffusion architecture from May 2025, which briefly appeared in preview before being withdrawn.

Performance metrics

The model demonstrates generation speeds exceeding 500 tokens per second. In testing on NVIDIA's NIM cloud API, DiffusionGemma 26B generated 2,409 tokens in 4.4 seconds when creating an image description. This represents a significant speed improvement over standard autoregressive language models.

Google's earlier Gemini Diffusion preview in May 2025 reportedly achieved 857 tokens per second, suggesting the architecture maintains high-speed generation capabilities.

Availability and access

The model is available as google/diffusiongemma-26B-A4B-it on Hugging Face. NVIDIA is currently hosting the model free of charge on their NIM cloud API platform, providing immediate access without local deployment requirements.

The 26B parameter model uses a diffusion-based approach to text generation rather than traditional autoregressive decoding, which enables parallel token generation and faster inference speeds.

Technical details

DiffusionGemma represents a departure from standard transformer architectures that generate tokens sequentially. Instead, the diffusion approach allows multiple tokens to be refined simultaneously during generation, similar to image diffusion models adapted for discrete text.

The "A4B" designation in the model name likely indicates architecture-specific configuration details, though Google has not released full technical specifications.

What this means

DiffusionGemma 26B validates diffusion architectures as a viable alternative to autoregressive generation for language models. The 500+ tokens/second speed, combined with Apache 2 licensing, makes this the fastest openly available language model by generation speed. This could shift inference economics for applications requiring high-throughput text generation, though quality comparisons with standard models like Llama or Gemma remain to be established through independent benchmarking.