Apr. 3 at 7:59 PM
Google announced Gemma 4, a family of open-source AI models released under an Apache 2.0 license. The company said developers have downloaded Gemma models over 400 million times since the first release, creating more than 100,000 variants.
Gemma 4 includes four model sizes: Effective 2B, Effective 4B, 26B Mixture of Experts, and 31B Dense. The 31B model ranks third globally on the Arena AI text leaderboard among open models, while the 26B ranks sixth. Google noted the models are built on the same research and technology as Gemini 3.
The models support advanced reasoning, agentic workflows with function calling and structured JSON output, code generation, and native multimodal processing (video, image, audio). Edge models offer a 128K context window, while larger models reach up to 256K, and all are trained on 140+ languages.
The unquantized bfloat16 weights of the 26B and 31B models fit on a single NVIDIA H100 (80GB GPU).
$GOOG $GOOGL $NVDA