Gemma 4 31B

Gemma 4 31B is the top dense model of the family Google DeepMind published on 11 March 2026. It has 30.7 billion parameters in 60 layers, a vocabulary of 262,000, a context window of 256,000 tokens and a vision encoder of roughly 550 million parameters; it takes text and images in and produces text out. Audio input, present in the smaller E2B, E4B and 12B models, is not offered here. The architecture is built around a trade-off between memory and awareness. Layers alternate between local sliding-window attention over 1,024 tokens and full global attention, with the last layer always global; the global layers use unified keys and values and Proportional RoPE to keep a quarter-million-token context affordable. Every model in the release is meant to reason, with thinking modes that can be configured rather than a separate reasoning variant, and the family adds native function calling and a native system role — the plumbing that decides whether a model can be built into software or only chatted with. The family spans five sizes, from E2B and E4B meant for phones to a 26B mixture-of-experts and this 31B dense model, and covers more than 140 languages. The commercially important change is legal, not technical: Gemma 4 is released under Apache 2.0, and the repository is not gated — anyone can fetch the weights without an account or an accepted set of terms. Gemma 3, ten months earlier, was still published under Google's own Gemma Terms of Use and required manual approval before download. Read on 25 August 2026, the 31B instruction-tuned repository records 8.79 million downloads in thirty days, more than any other frontier-class open-weight model in the catalogue.

▸News

Open weights stopped asking permission — of the forty most downloaded models, only Meta's still need an account

8/25/2026 · Hugging Face (model repositories and licence files)

▸Videos

No videos yet.

▸Reviews