Gemma 4 31B IT

Gemma 4 31B IT sits within the Gemma 4 family developed by Google DeepMind as a multimodal, open-weights model that processes text and image inputs while generating text output. It ships in both pre-trained and instruction-tuned variants under the Apache 2.0 license, joining four other Gemma 4 sizes that span efficient and high-capacity tiers. The 31B variant is positioned alongside a 26B A4B Mixture-of-Experts model, giving developers a choice between a dense 31B checkpoint and MoE alternatives for different deployment profiles. Across the family, Gemma 4 is presented as well-suited to text generation, coding, and reasoning tasks, with the instruction-tuned "IT" variant specifically optimized for assistant-style interaction.

The model is described as a capable reasoner with configurable thinking modes, making it a practical fit for agentic workflows that combine step-by-step deliberation with tool calling and structured assistant behavior. Its multimodal support for variable-resolution images broadens use cases into document understanding, visual question answering, and chart or diagram interpretation alongside pure text work. The five-size Gemma 4 lineup targets deployment ranging from high-end phones up through laptops and servers, and the 31B IT checkpoint is the most substantial dense option for server-class hardware. For teams that need a balance of reasoning quality, multimodal flexibility, and the freedom to self-host with permissive licensing, this checkpoint represents the top-end of the dense Gemma 4 tier.