Google has introduced a new family of open models of artificial intelligence - Gemma 4. According to the company, these are the most intelligent open models to date, created on the basis of chat-based technologies.Gemini 3 bot. Since the first generation, developers have downloaded Gemma more than 400 million times and created over 100,000 variants in the Gemmaverse ecosystem.

Four versions were available: the compact multimodal Effective 2B and Effective 4B (2.3 billion and 4.5 billion parameters), as well as the more powerful 26B MoE and the flagship 31B Dense (26 and 31 billion parameters). Junior models can be run on smartphones and laptops, older ones require Nvidia H100-level graphics accelerators. In the Arena AI ranking, the flagship version ranks third among open-text models.

Gemma 4 supports complex reasoning, function invocation, code generation, video and audio processing, and up to 256,000 context tokens. Models are available in Google AI Studio, Google AI Edge Gallery and through third-party frameworks (Hugging Face, Ollama, Nvidia NIM and others).