Gemma 4 31B IT is the instruction-tuned flagship of Google DeepMind's open Gemma 4 family, distilled from Gemini 3 research and built around a 30.7-billion parameter dense architecture. It is released under Apache 2.0, paired with a pre-trained sibling, and joins sibling variants that span dense and mixture-of-experts designs from E2B up through 26B A4B and 31B. The family targets text generation, coding, and reasoning workloads across deployment environments ranging from phones to servers, with the 31B Instruct variant tuned for chat, task completion, and agentic workflows.

The model processes text and image inputs into text outputs, drawing on a 256K-token context window with hybrid attention for efficient long-context handling, and accepts video on selected variants. It supports a configurable thinking mode, native function calling, structured output, and multilingual understanding across more than 140 languages. Reported strong suits include reasoning benchmarks such as GPQA Diamond at 85.7%, alongside coding and document understanding tasks, making it well suited for coding assistants, multimodal analysis, document Q&A, and tool-driven agents that benefit from an open-weights deployment path.