Gemma 4 31B

Gemma 4 31B is Google DeepMind's 30.7 billion parameter dense multimodal model — not a mixture-of-experts design — accepting text and image input and returning text. It offers a 262,144-token context with a 32,768-token maximum output, a configurable thinking mode, native function calling, and multilingual coverage across more than 140 languages, all under an Apache 2.0 license. Its dense architecture, including a roughly 550M-parameter vision encoder, makes it the most hardware-frugal model in the library while still competing with far larger frontier systems.