Gemma 4
Google's open-source reasoning & agent model, Apache 2.0, four sizes
What is Gemma 4
Gemma 4 is Google DeepMind's open model family released April 2026, built for reasoning and agentic workflows. Four sizes: 31B (Dense, strongest), 26B-A4B (MoE, efficient), E4B and E2B (lightweight). Apache 2.0 license, no commercial restrictions.
How to Run Locally
Easiest way: ollama run gemma4:31b. The 31B version needs ~20GB VRAM/RAM, 26B-A4B only ~4GB active params. Supports speculative decoding for 29% speedup (50% on code).
Coding Benchmark Comparison
Gemma 4 31B excels at reasoning and agent tasks, but slightly behind Qwen 3.6-35B-A3B on coding benchmarks (64.7% vs 71.4%). Advantages: most permissive Apache 2.0 license, best native function calling support.