Google DeepMind · 2026-04-02 · major
Gemma 4 — open models for reasoning and agents
Google DeepMind ships four Gemma 4 sizes (E2B, E4B, 26B MoE, 31B dense) under Apache 2.0 with 256K context, native vision/audio, and agentic tool-use built in. The 31B scores 89.2% on AIME 2026.
Four open models from Google under Apache 2.0 that bring frontier-tier reasoning and native tool-use to consumer GPUs and phones.
Key specs
| License | Apache 2.0 |
|---|---|
| GPQA | 84.3% (31B) |
| Variants | E2B, E4B, 26B MoE, 31B dense |
| Aime 2026 | 89.2% (31B) |
What is it?
Gemma 4 is Google DeepMind's latest family of open-weights models, built from the same research as Gemini 3 and released on April 2, 2026. It ships in four sizes: E2B (2.3B effective) and E4B (4.5B effective) for edge and mobile, a 26B Mixture-of-Experts (4B activated) for mid-range hardware, and a 31B dense model for workstations. Every variant ships under Apache 2.0 with no usage caps.
How does it work?
All four models are natively multimodal (text, image, video; audio on the small models) with context windows up to 256K tokens. The 31B instruction-tuned model scores 89.2% on AIME 2026 (up from 20.8% on Gemma 3 27B), 84.3% on GPQA Diamond, and 86.4% on the tau2-bench agentic tool-use benchmark. Native function calling is built in, so the models can drive tools and APIs without prompt hacking.
Why does it matter?
This is the first Apache 2.0 model family that competes with frontier proprietary models on both reasoning and agentic benchmarks while still fitting on a phone (E2B) or a single GPU (31B). The license means anyone can deploy, modify, and redistribute without restrictions or revenue caps.
Who is it for?
On-device ML engineers, self-hosters, agent framework authors, Android developers.
Try it
huggingface.co/google/gemma-4-31B-it