Gemma 4 31B It

Overview. Gemma 4 31B IT is Google DeepMind's open-weights, instruction-tuned dense model with 30.7B parameters, released in April 2026 under the Apache-2.0 licence. It accepts text, image, and video input and returns text output, backed by a 262,144-token context window. Developers reach for it when they need an open-weights multimodal model they can self-host or run through Inference without licensing friction.

Key strengths. The model's coding index (43.4 on Artificial Analysis) is its clearest strength, ahead of its general intelligence index (29.7), suggesting more reliable performance on code tasks than on broad reasoning benchmarks. Its 262K-token context window supports long-document and long-conversation workloads well beyond typical chat-model limits. Chain-of-thought reasoning is available as an opt-in mode, letting a developer trade latency for deeper reasoning only when a task calls for it, rather than paying that cost on every request.

When to use it. Use this model for long-context tasks — large document analysis, multi-turn agents with extended history, or codebases that exceed shorter context windows — where its 262K window is the deciding factor. It's also a fit for multimodal tasks that combine text with image or video input. Because its general intelligence index sits in the mid range and it has no confirmed domain-specific benchmarks for math, legal, or finance, it's better suited to medium-complexity tasks than to specialised high-stakes reasoning.

Limitations. Independent benchmark coverage for math, legal, medical, and finance domains is not currently available, so quality in those areas is unverified rather than confirmed strong. Its general intelligence index trails frontier open models, and its opt-in thinking mode adds latency when enabled rather than running by default. Training data has a cutoff around January 2025, so it won't reflect more recent events or releases.

Sovereignty note. Gemma 4 31B IT is built by Google, a US-based creator with CLOUD Act exposure, so a hosted endpoint of this model carries that exposure. Because the weights are open under Apache-2.0, teams that need full data residency control can self-host the model instead of relying on a third-party endpoint.