Google Gemma 4 unter Apache-2.0-Lizenz: 31B-Modell belegt Platz 3 im LM Arena

Google Gemma 4 Launches Under Apache 2.0: 31B Ranks #3 on LM Arena

Google DeepMind has released the Gemma 4 family under the Apache 2.0 license — a significant licensing upgrade from its prior bespoke terms. Four model sizes ship: E2B and E4B for on-device use (multimodal text, vision, and audio); a 26B mixture-of-experts (128 experts, 8 active per pass — the first MoE in the Gemma family); and a 31B dense model ranking third globally on the LM Arena leaderboard, outperforming models 20× its size. All large models support 256K context, native function calling, and structured JSON output. On-device models ship a 35M-parameter audio conformer and a variable-resolution vision encoder. A Gemma 4 E2B browser agent with 100% local WebGPU execution also launched simultaneously.

Why It Matters

Apache 2.0 makes Gemma 4 infrastructure-ready for production deployments at scale, and LM Arena #3 at 31B is a genuine frontier result at an accessible parameter count — narrowing the gap between proprietary and open-weight models for agentic workloads.