Gemma 4 31B: frontier reasoning over a quarter-million tokens, and where to run it

Apache 2.0 from Google, weights free, no terms to accept, which for a model at this capability level is the notable part. Self-hosting 30.7 billion parameters means a data-centre GPU and either the Hugging Face transformers library or a serving framework like vLLM. In your browser it runs on CNAPS Studio, included in the basic plan.

What you get is the reasoning end of the open-weight field. It scores 89.2 out of 100 on this year's AIME competition mathematics and 84.3 on graduate-level science questions, numbers that until recently belonged to closed models only. It reads text, images and video, calls external tools natively, and holds 256,000 tokens of context, which is roughly a 200,000 word document in one pass.

The cost is that it cannot hear and it is not quick. Audio input, which its small sibling handles, is absent here. And thinking mode is on by default: for hard problems Google suggests allowing up to 81,920 tokens of output so it has room to work, which is a long wait. Three multimodal models sit in the CNAPS Studio catalog and this is the one to bring to a hard problem.

What it actually is

An instruction-tuned multimodal transformer, the largest in the Gemma 4 family. It reads text, images and video, writes text, and its native function calling means it can be wired to external tools and act across several steps rather than answering once.

The context window is the specification that changes what you can attempt. At 256,000 tokens you can put an entire codebase, a full contract set, or a year of conversation history in front of it without chunking, and Google publishes a long-context retrieval score for exactly that case.

Settings in CNAPS Studio follow Google's guidance and split by task. Max new tokens should be 32,768 for most work and up to 81,920 for competition maths or code, because thinking consumes output budget. Temperature should be 0.6 for precise coding and 1.0 for open-ended work. Top-P is 0.95 with thinking on and 0.8 in standard mode, Top-K is 20 throughout, and the thinking toggle is on by default.

Parameters30.7 billionContext window256,000 tokens, roughly 200,000 wordsModalitiesText and image in, text out; video read as framesAudioNot supported on this sizeThinking modeBuilt in, on by defaultFunction callingNativeLanguages35 or more out of the box, pre-trained on 140 or moreLicenseApache 2.0Downloads last month7,973,406

The published numbers

Google publishes a wide scorecard. As percentages out of 100: 85.2 on MMLU Pro for broad expert knowledge, 89.2 on AIME 2026 for competition mathematics without tool use, 80.0 on LiveCodeBench v6 for real coding problems, 84.3 on GPQA Diamond for graduate science, 76.9 on MMMU Pro for reading images and text together, and 85.6 on MATH-Vision for maths presented as pictures.

Two figures use their own scales. The Codeforces rating of 2,150 is a competitive programming ELO rather than a percentage, and it puts the model in the range of a strong human competitor. And the long-context retrieval test, which hides eight facts inside a 128,000 token document and asks for them back, returns 66.4, which is the honest reminder that a very large context window is not the same as perfect recall across it.

Everything here is vendor-reported. What makes that less worrying than usual is that these are public benchmarks with public test sets, re-run constantly by people with no stake in Google, and the model is downloaded nearly eight million times a month. Claims made in that much daylight get corrected quickly if they are wrong.

How it compares to the other multimodal models here

Gemma 4 31BGemma 4 E2BQwen3.6-35B-A3BSize30.7B, needs a data-centre GPU2.3B effective, runs on a phone or laptopSee its own postContext256,000 tokens128,000 tokensSee its own postAudio inputNoYes, up to 30 secondsSee its own postMMLU Pro85.260.0See its own postAIME 202689.237.5See its own postGPQA Diamond84.343.4See its own postLicenseApache 2.0Apache 2.0See its own post

The two Gemma columns are genuinely comparable, since the same publisher ran the same benchmarks and reported them together. That is unusual in this series and worth using. The gaps are wide, and widest exactly where you would expect: competition mathematics separates them by more than fifty points, general knowledge by twenty-five.

Reach for Gemma 4 31B when the problem is genuinely hard, when the document will not fit anywhere else, or when you want an agent that can call tools and follow a plan. Reach for Gemma 4 E2B when the input includes audio, when the task is routine, or when speed and cost matter more than the last twenty points of accuracy. Reach for Qwen3.6-35B-A3B for a second opinion from a different publisher at a similar scale.

What to chain it with

It takes text and images and returns text, which puts it at the end of most chains rather than the middle. Its job is usually to make sense of whatever the models before it produced.

Two chains use it well. For documents, DeepSeekOCR or GLM-OCR turns scans into markdown and this model reads the lot at once, which is where the 256,000 token window pays for itself: a hundred-page contract set goes in whole rather than in pieces that lose cross-references. For images, a detector such as RF-DETR Medium or a captioner such as BLIP (Image Description) produces structured findings across a batch, and this model turns them into a written summary or a judgment.

Open Gemma 4 31B in CNAPS Studio, give it a long document with thinking on and again with it off, and see what the extra reasoning time actually buys on your material.

Sources

• huggingface.co/google/gemma-4-31B-it for the Apache 2.0 license, the 30.7 billion parameter count, the 256,000 token context, the supported modalities, the absence of audio at this size, every benchmark score quoted above, and the monthly download count.
• huggingface.co/google/gemma-4-E2B-it for the comparison column.
• docs.cnaps.ai/ai-model-hub for the CNAPS Studio settings and Google's recommended values for each task type.