Google Launches Gemma4 Open Source Large Model with Record-breaking Per-parameter Intelligence Level

Google officially launched the Gemma4 open source large model, said to be its most intelligent open source model to date. Designed for advanced reasoning and agent workflows, it offers four specifications: Efficient 20B (E2B), Efficient 40B (E4B), 26B Mixture of Experts (MoE), and 31B dense model, all capable of handling complex logic and agent workflows.

Gemma4 excels in performance: the 31B model ranks third among global open source models on the Arena AI text list, and the 26B model ranks sixth—even outperforming models 20 times its size. This means developers can achieve cutting-edge AI capabilities with less hardware cost.

Additionally, Gemma4 has core advantages like advanced reasoning, native agent workflow support, high-quality code generation, multi-modal processing (vision/audio), up to 256K context window, and support for over 140 languages. Edge models can also run offline on devices like mobile phones.