Google’s DiffusionGemma: 1,000 tokens per second and open weights

Google’s DiffusionGemma reaches 1,000 tokens per second on an NVIDIA H100 and records over 700 tokens per second on an NVIDIA GeForce RTX 5090. The model is four times faster than regular Gemma and is described as a speed-focused variant rather than a quality upgrade. DiffusionGemma is released under the Apache 2.0 license, with model weights available on Hugging Face.

Google reported that DiffusionGemma achieves over 700 tokens per second on an NVIDIA GeForce RTX 5090. The company included the speed figure in its announcement of the model. Google said, “This is a speed model, not a quality upgrade.”

In comparisons with standard Gemma 4, DiffusionGemma trails on output quality. Google’s materials state that DiffusionGemma prioritizes throughput over matching Gemma 4’s output quality. The announcement frames the release as a trade-off between higher throughput and lower output quality relative to standard Gemma 4.

These are the performance claims and model-comparison statements provided by Google. The company emphasizes DiffusionGemma’s speed orientation rather than parity with Gemma 4 on output quality.

DiffusionGemma is a diffusion-based text model that begins generation from a canvas of random placeholder tokens and then iteratively locks tokens in place across multiple steps. The model processes blocks of tokens per forward pass, with one reported configuration using 256 tokens per forward pass. During these forward passes the GPU remains actively utilized to sustain throughput. The iterative locking process progresses until tokens are fixed and output is produced.

The model uses bidirectional attention during generation, which allows each token being generated to see every other token in the sequence while generation proceeds. Efficiently running DiffusionGemma requires a drafter module, described as a lightweight component that proposes token blocks in parallel. That drafter operates using speculative decoding to suggest blocks of tokens concurrently with the diffusion steps. Speculative decoding and parallel token proposals are presented as operational requirements to keep the model’s computation efficient.

The preceding paragraphs describe DiffusionGemma’s technical mechanisms and operational requirements. They do not repeat the model’s performance or quality comparisons presented elsewhere.

Inception Labs shipped Mercury 2 in February 2026 as the first commercial diffusion reasoning model, and the company claimed its speeds are five times faster than speed-optimized competitors. The announcement describes Mercury 2 as a commercial product in the diffusion reasoning category. The DFlash framework is described as enabling over 6x speedup on some tasks. The source lists these developments alongside other diffusion-focused work.

The source identifies DiffusionGemma as the first major open release from a tier-one lab. That characterization distinguishes DiffusionGemma’s open availability from commercial releases noted in the source. The source frames DFlash and Mercury 2 as examples of ongoing efforts targeting throughput or commercial deployment. Those entries are presented as part of the same landscape described in the source.

Together, these items sketch the contemporaneous models and tooling mentioned in the source. They are listed as context for DiffusionGemma in the source material.

DiffusionGemma’s release is the first major open release from a tier‑one lab and makes the model and related artifacts openly available. The model implements diffusion-based generation techniques that begin from a canvas of random placeholder tokens, iteratively lock tokens, use bidirectional attention during generation, and rely on a drafter module that proposes token blocks via speculative decoding. The release presents throughput-oriented diffusion techniques alongside open artifacts for further use.