Google DeepMind's DiffusionGemma hits 1,100 tokens/sec on H100
Open-weights multimodal model with 256K context and MoE architecture
Google DeepMind has open-sourced DiffusionGemma 26B A4B IT, a multimodal generative model that accepts text, image, and video inputs to produce text via discrete diffusion. Built on the Gemma 4 MoE architecture, it packs 25.2 billion total parameters but activates only 3.8 billion per token, making it extremely efficient. The encoder-decoder design with bidirectional attention generates tokens in parallel 256-token blocks, enabling inference speeds exceeding 1,100 tokens per second at low batch sizes on NVIDIA Hopper H100 hardware.
The model delivers a 256-thousand token context window, configurable thinking modes for step-by-step reasoning, native function calling for agent workflows, and multilingual support across 35+ languages. NVIDIA released a quantized version (NVFP4) using Model Optimizer, available for both commercial and non-commercial use. Developers can deploy it for conversational AI, code generation, document OCR, chart comprehension, PDF parsing, video analysis, and agentic tasks — all with minimal latency.
- MoE architecture: 25.2B total / 3.8B active parameters for efficient inference
- 1,100+ tokens/sec generation on H100 (FP8) using parallel 256-token block decoding
- 256K context window, native function calling, and 35+ language support
Why It Matters
A high-speed, open‑source multimodal model enables low‑cost AI agents and real‑time video understanding.