Google DeepMind released DiffusionGemma on June 10, 2026, a new text-generation model that produces text in parallel blocks rather than sequentially.
The company says it reaches up to 1,000 tokens per second on Nvidia GPU hardware.
According to a report, DeepMind's benchmarks show DiffusionGemma runs 4x faster than previous Gemma autoregressive models on equivalent compute. A separate benchmark report confirmed 10x higher token throughput in long-context inference tests conducted on Nvidia hardware.
How DiffusionGemma Works
Standard large language models generate one token at a time. DiffusionGemma generates entire text blocks simultaneously using a diffusion-based architecture. The approach reduces latency sharply for long outputs. DeepMind states the model self-corrects complex markdown and structured formats during generation.
That capability is targeted at developers building code assistants, documentation tools, and structured data pipelines. The model is optimized for local deployment on Nvidia RTX consumer GPUs and DGX enterprise systems.
Also Read: SpaceX’s $75B IPO May Be In Trouble As Warren Pushes SEC Delay
Background
Google DeepMind has released several Gemma variants over the past year, each expanding the open-weights model family for different use cases. DiffusionGemma marks the first time DeepMind has applied a diffusion architecture to text generation within the Gemma line.
Prior diffusion text models from other labs have shown speed advantages in research settings but limited real-world deployment. DeepMind's release brings the approach to a widely used model family with existing developer tooling.
The timing follows Anthropic's release of Claude Fable 5 earlier this week, which set new benchmarks on reasoning and coding tasks. DeepMind's focus on raw inference speed at the hardware level targets a different competitive dimension, prioritizing throughput for high-volume deployment rather than benchmark scores.
Nvidia benefits directly. The DGX and RTX optimization cements Nvidia hardware as the default platform for frontier model inference at the local level.
What to watch is developer adoption speed and whether DiffusionGemma's throughput figures hold across non-Nvidia hardware configurations.
Read Next: SpaceX's $250B IPO Is Draining Crypto Liquidity, Traders Fear