Google Gemma 4 31B Instruct

• 🧠 Dense 30.7B open model from Google DeepMind for reasoning
• 🆕 Configurable thinking modes toggled via a reasoning token
• 📏 256K-token context window for long documents and code
• 👁️ Handles text and image input; video processed as frames
• 🔧 Native function calling for agentic, tool-using workflows
• 🏢 Quantized checkpoints target consumer GPUs and workstations
• 🔒 Apache 2.0 license; open pre-trained and instruction-tuned weights
• 📚 Hybrid local/global attention with Proportional RoPE for long context

Google

Google is an American multinational technology corporation and one of the world's most valuable brands. A subsidiary of parent company Alphabet Inc., Google operates across search, cloud computing, consumer electronics, and artificial intelligence. Its…

Gemma 4 31B Instruct is the dense flagship of Google DeepMind's Gemma 4 family, a 30.7B-parameter multimodal model that accepts text and image input (and can process video as sequences of frames) while generating text output. It offers a 256K-token context window, native function calling, and configurable thinking modes, aimed at running reasoning, coding, and multimodal tasks under an Apache 2.0 license.

Architecturally it is a dense transformer paired with a vision encoder, using a hybrid attention scheme that interleaves local sliding-window layers with full global attention and Proportional RoPE (p-RoPE) for efficient long-context handling; quantization-aware and w4a16 checkpoints are published for smaller-footprint deployment.

Relative to the sibling Gemma 4 26B A4B Instruct, a Mixture-of-Experts variant with fewer active parameters, this 31B is dense—trading that inference efficiency for the family's highest-quality tier. Against the previous generation Gemma 3 27B, Google DeepMind highlights Gemma 4's built-in reasoning with configurable thinking, native system-prompt and function-calling support, and coding improvements.

Google DeepMind publishes instruction-tuned results in the official Gemma 4 31B model card, spanning reasoning, coding, vision, long-context, and safety tasks, and states the models undergo the same safety evaluations as its proprietary Gemini models.