جیما 4 31B

An open model from Google DeepMind, built from Gemini 3 research to maximise intelligence per parameter. It is the dense 31B member of the Gemma 4 family, which Google describes as maximising raw quality and providing a powerful foundation for fine-tuning, and it handles text and image input across a 256K context. Google reports it outcompeting models twenty times its size.

Gemma is the open half of Google's model line, and Gemma 4 is released under the Apache 2.0 licence - the weights are published, and anyone may download, run, fine-tune and ship them commercially. In practice that means two routes to the same model: run it yourself on hardware you control, with nothing leaving your machine, or call it here and skip the hardware. Google sizes the family for the first route explicitly, and says the unquantized weights of this 31B fit on a single 80GB accelerator, with quantized builds running on consumer GPUs. That is a claim about where it fits, not a promise that it is easy to operate; serving it well is still work.

Inside the family the 31B is the dense member, the one Google positions for raw quality and as the base to fine-tune from, with the sparse 26B beside it for latency and two much smaller models for phones and edge devices. The lab's own comparison is against Gemma 3 27B, and the margins are large: 85.2% against 67.6% on the multilingual MMMLU, 76.9% against 49.7% on MMMU Pro, 89.2% against 20.8% on AIME 2026, 80.0% against 29.1% on LiveCodeBench v6, 84.3% against 42.4% on GPQA Diamond, and 86.4% against 6.6% on the retail split of tau2-bench. All of those are with thinking enabled.

The feature list is aimed at building rather than chatting. Function calling, structured JSON output and native system instructions are supported directly, so an agent can be wired to real tools without a wrapper coaxing the format out of it. Image and video input is native across the family at variable resolutions, with optical character recognition and chart reading called out specifically, and Google trained the models on more than 140 languages.

The context window is 256K on this size, which is generous for an open model and short of the million tokens the Gemini line carries. Audio input belongs to the two edge models and not to this one. And the claim Google is making throughout is intelligence per parameter, not parity with a frontier model: this is the most capable thing of its size, which is a different statement from the most capable thing.