Gemma 4 31B Instruct

Gemma 4 31B Instruct is a dense instruction-tuned model from Google DeepMind that continues the Gemma family's line of open releases under the Apache 2.0 license. It is built around a roughly 30.7 billion parameter architecture and accepts images and text as input while producing text outputs, making it a multimodal generalist rather than a text-only assistant. Within the Gemma 4 generation it sits alongside sibling variants such as Gemma 4 26B A4B, giving teams a choice between a dense 31B model and a mixture-of-experts alternative at a similar scale, which is useful when comparing quality-per-parameter against deployment cost.

Practically, the model is aimed at builders who want transparent weights and flexible behavior for reasoning, tool calling, and structured output workflows. Independent catalog benchmarks place it at a coding index of 43.4, an agentic index of 14.4, and an intelligence index of 29.7, positioning it as a capable mid-size option rather than a frontier leader. On Berget's sovereign infrastructure it runs as an OpenAI-compatible endpoint hosted entirely in Europe, which suits organizations that prioritize data residency, open-source lineage, and integration with agentic coding or document-analysis pipelines without depending on closed providers.