google-deepmind/gemma

Gemma 4 is the most capable open model family Google DeepMind has ever released. Four variants, all under Apache 2.0. The flagship 31B dense model ranks #3 on Arena AI at 1,452 Elo, outperforming models twenty times its size. Released April 2, 2026. The model lineup: E2B (2.3B effective parameters), E4B (4.5B effective), 26B MoE (3.8B active / 26B total), and 31B Dense. All four handle text and image inputs natively. E2B and E4B also support audio input, making them the smallest multimodal models with audio understanding available as open weights. Benchmark numbers for the 31B dense model: 85.2% on MMLU Pro, 89.2% on AIME 2026 (mathematical reasoning), 80.0% on LiveCodeBench v6, Codeforces Elo of 2,150, 76.9% on MMMU Pro (vision), and 85.6% on MATH-Vision. The 26B MoE variant ranks #6 on Arena AI with only 3.8B active parameters, making it brutally efficient for inference. The Apache 2.0 license is the headline change from previous Gemma releases. No monthly active user limits. No acceptable-use policy enforcement. No restrictions on commercial deployment, sovereign AI initiatives, or redistribution. This puts Gemma 4 on the same licensing footing as Qwen 3.5 and more permissive than Llama 4's community license. Hardware support spans NVIDIA GPUs, AMD ROCm, Google TPUs, and even Raspberry Pi for the E2B variant. Available on Hugging Face, Kaggle, Ollama, and Google AI Studio. The E2B model running on a Raspberry Pi opens on-device AI for embedded and edge use cases that were previously impossible with open weights. Built from the same research behind Gemini 3, with purpose-built optimizations for reasoning and agentic workflows. Native tool use, structured output generation, and multi-turn planning make Gemma 4 viable for agent systems that previously required proprietary API models. If you run local models, Gemma 4 is now the benchmark to beat. For local inference tooling, check Ollama which supports Gemma 4 out of the box. For AI coding agents that benefit from local models, explore oh-my-codex. Related reading: AI code editor rankings showing how open models compare in coding tasks.