Google Gemma 4 31B Instruct

• 🆕 Dense 31B open-weights model from Google DeepMind, Apache 2.0.
• 📏 256K-token context with hybrid local/global attention and p-RoPE.
• 👁️ Multimodal: text, image, and video as frame sequences.
• 🧠 Configurable thinking modes for step-by-step reasoning.
• 🔧 Native function calling and structured output for agentic workflows.
• 🌐 Pre-trained across 140+ languages.
• 🏢 Targets consumer GPUs and workstations.

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, maximum-quality member of Google DeepMind's open Gemma 4 family, built for consumer GPUs and workstations rather than edge devices. It handles text and image inputs, processes video as sequences of frames, and generates text, with a 256K-token context window and support for over 140 languages under the Apache 2.0 license.

Against its same-family predecessor Google Gemma 3 27B Instruct, Gemma 4 introduces several documented changes: configurable thinking modes that emit internal reasoning before a final answer, and native function calling for agentic workflows.

Architecturally, Gemma 4 uses a hybrid attention mechanism interleaving local sliding-window and full global attention, with unified Keys and Values in global layers and Proportional RoPE to aid long-context performance. It sits alongside the latency-focused Google Gemma 4 26B A4B Instruct, a Mixture-of-Experts sibling that activates only a subset of its parameters per token for faster inference, whereas the 31B Dense model keeps all parameters active for quality.

Both pre-trained and instruction-tuned variants are released as open weights, and the model cards note that Gemma 4 underwent safety evaluations and sensitive-data filtering during training.