Google's open-weight language models family, with 4 sizes from 2B to 31B parameters. The 31B model scores 89% on AIME 2026 math and supports text, image, audio and video inputs.
Released 2 April 2026 under Apache 2.0 license with models from 2B to 31B parameters.
Updated 2 Apr 2026Checked 10 Oct0 updates this week
Gemma 4 trains on text, images, audio and video. The larger models combine a base dense model with mixture-of-experts for efficiency.
All sizes understand multiple input types. The models include function calling and can output structured JSON, enabling agentic behavior.
Weights are published fully, with no restrictions on commercial fine-tuning or redistribution. Users can run models privately on their own hardware.
| Spec | Gemma 4 |
|---|---|
| Available model sizes | 2.3B, 4.5B, 26B MoE, 31BR (reported) |
| AIME 2026 score (31B) | 89.2%R (reported) |
| LiveCodeBench score (31B) | 80.0%R (reported) |
| MMLU Pro score (31B) | 85.2%R (reported) |
| Arena AI leaderboard rank (31B) | 3rd globallyR (reported) |
R reported by the company
Google DeepMind
Gemini, AlphaFold, Genie, Veo, TPU chips
Google DeepMind is Alphabet's AI lab, formed in 2023 from DeepMind and Google Brain. It builds the Gemini models and leads in AI for science. AlphaFold won a share of the 2024 Nobel Prize in Chemistry.
