Gemma 4 12B

Evidence (4)

• 7Simon Willison·6/10/2026model_release
  Google Releases Open-Weight DiffusionGemma Model

  Google has released DiffusionGemma, an open-weight Gemma model under the Apache 2 license, reviving work previously seen in an experimental Gemini Diffusion preview. NVIDIA is currently hosting the model for free via its NIM cloud API, and early use suggests it can generate output at high speed. The release highlights continued interest in diffusion-based language models and their performance potential.

• 6Google DeepMind Blog·6/9/2026model_release
  Google DeepMind unveils Gemma 4 12B

  Google DeepMind introduced Gemma 4 12B, a unified multimodal model designed without a separate encoder component. The announcement highlights a model architecture aimed at handling multiple input types in a single system. No additional technical details were provided in the headline text.

• 7InfoQ·6/8/2026model_release
  Gemma 4 12B Brings Multimodal On-Device Agents to Laptops

  Google says Gemma 4 12B is designed to deliver agentic, multimodal intelligence directly on a laptop and can be paired with Google AI Edge for local experimentation. The model is positioned for tasks such as autonomous data processing, visual insight generation, webpage building, and tool execution.

• 8Hugging Face Blog·4/2/2026model_release
  Gemma 4 brings on-device multimodal AI to Hugging Face

  Hugging Face announced Gemma 4, a new frontier multimodal model designed to run on-device. The release highlights compact, capable AI that can process multiple input types while keeping inference local for lower latency and improved privacy.