Google Gemma 4 Released: 31B Open Model Ranks #3 Globally

Google has released Gemma 4, the latest generation of its open-weight model family, in four variants ranging from 2.3 billion to 31 billion parameters.

Gemma 4 31B: #3 globally

The flagship Gemma 4 31B Dense model ranks third globally on the Arena AI leaderboard among open-weight models — behind only much larger models. This makes it the strongest open model Google has ever released and a serious competitor to Meta's Llama family.

Model variants

The four Gemma 4 variants — 2.3B, 7B, 14B and 31B — are designed to cover different deployment scenarios, from on-device mobile applications to cloud inference. All variants support a 128K token context window and are optimized for Google's TPU hardware as well as standard NVIDIA GPUs.

Multimodal capabilities

For the first time in the Gemma series, the 14B and 31B variants include native vision capabilities, allowing them to process images alongside text. This brings Gemma closer to feature parity with closed multimodal models.

All Gemma 4 weights are available on Hugging Face and Google AI Studio under a permissive open license allowing commercial use.