Google DeepMind drops DiffusionGemma, Gemma 4 12B, and new AI safety research
DiffusionGemma delivers 4x faster text generation while Gemma 4 12B goes multimodal...
Google DeepMind unveiled a series of June 2026 updates spanning AI safety and new models. The company announced increased investment in multi-agent AI safety research, signaling a focus on aligning systems that can interact and coordinate. Alongside that, two new models launched: DiffusionGemma, which achieves four times faster text generation compared to its predecessor, and Gemma 4 12B, described as a unified, encoder-free multimodal model that can handle vision and language tasks without separate encoders. Additionally, Gemini 3.5 Live Translate was introduced, providing fluid, natural voice translation for real-time conversations. These updates collectively push Google DeepMind's portfolio toward faster inference, broader modality support, and safer multi-agent deployments. The research and model releases aim to give developers more efficient tools while addressing growing concerns about autonomous AI systems.
- Google DeepMind invested more in multi-agent AI safety research, focusing on coordination and alignment.
- DiffusionGemma offers 4x faster text generation than previous versions, improving inference speed.
- Gemma 4 12B is a unified encoder-free multimodal model, handling vision and language without separate encoders.
Why It Matters
Faster, safer, multimodal models and real-time translation expand practical AI use for enterprises and developers.