Gemma 4 31B: The Workstation-Sized Member of Google's Open Family

What Gemma 4 31B is

Gemma 4 31B is the largest dense model in Google DeepMind's Gemma 4 family, released under Apache 2.0 with open weights in both pre-trained and instruction-tuned variants. It is a 30.7 billion parameter dense model — not mixture-of-experts — across 60 layers, with a 256K token context window (recorded here as 262.1K, the same number written out) and multilingual support across more than 140 languages. Google position the family from mobile and edge deployment up through consumer GPUs and workstations, and 31B is the workstation end of that range: the model you run when you have a serious GPU and want the most Gemma you can fit on it.

Where it sits in the family, and what it gives up

Google publish the whole family's specifications side by side. E2B and E4B are the edge models at 2.3 and 4.5 billion effective parameters with a 128K context; 12B Unified and the 26B A4B mixture-of-experts model (25.2 billion total, 3.8 billion active, 128 experts with 8 active and 1 shared) sit in between with 256K. All of them share a 262K vocabulary. E2B, E4B and 12B handle text, image and audio, while 31B is text and image only — it has a roughly 550M parameter vision encoder and no audio encoder at all.

One migration detail from Gemma 3

Gemma 4 moves to the standard system, assistant and user roles, where Gemma 3 used its own convention. Anything written against the older chat template needs updating. Google also document control tokens for managing the thinking process, which is how reasoning is switched on and bounded.

LMArena

On LMArena's text leaderboard Gemma 4 31B has 1452 Elo (rank 77); by category, 1502 Elo in Coding (rank 75), 1469 Elo in Math (rank 55), 1498 Elo in Expert (rank 46) and 1420 Elo in Creative Writing (rank 86).

Licence and price

Apache 2.0 with open weights, a 256K context, image input and more than 140 languages. Through third-party providers it runs at around $0 per million input tokens, an aggregated figure rather than a first-party one.

Architecture and capabilities

Gemma 4 uses a hybrid attention mechanism that interleaves local sliding-window attention with full global attention, with the final layer always global; on 31B the sliding window is 1,024 tokens. Global layers use unified keys and values and Proportional RoPE to reduce memory at long context. All Gemma 4 models have configurable thinking modes, native function calling and native support for the system role. Image input is accepted at variable aspect ratios and resolutions, covering object detection, document and PDF parsing, screen and UI understanding, chart comprehension, OCR and handwriting; video is analysed as frame sequences. Google state out-of-the-box support for more than 35 languages, with pre-training on more than 140.

Google’s benchmark table

For the instruction-tuned 31B, Google report MMLU Pro 85.2%, AIME 2026 without tools 89.2%, LiveCodeBench v6 80.0%, a Codeforces rating of 2150, GPQA Diamond 84.3%, MMMU Pro 76.9% and 66.4% on the 128K MRCR v2 8-needle test, against Gemma 3 27B's 67.6%, 20.8%, 29.1%, 110, 42.4%, 49.7% and 13.5%. These are Google's own figures.

Official Resources