Gemma 4 31B
Made by Google DeepMind. Epoch AI does not state how it can be accessed. No Top tier or Strong results yet among the models listed here.
In short
What it does well
No Top tier or Strong ratings on the four main tests yet.
Where it falls short
- Coding: Behind the leaders (60th of 77 listed models)
- Writing: Behind the leaders (59th of 77 listed models)
What we don't know yet
- Hard maths problems: Not enough data
- School-competition maths: Not enough data
- Graduate-level science questions: Not enough data
- Short factual questions: Not enough data
- Fixing real software bugs: Not enough data
- Difficult coding tasks: Not enough data
The four ratings
- Top tier
- Strong
- Middle
- Behind the leaders
- Not enough data
Ratings compare only the 77 models listed here, in quarters, best first. “Not enough data” means no result yet. It is never a low score.
- Writing
- Behind the leadersHow often people preferred its answers to writing prompts. 1,416 on Arena's scale · Source for Gemma 4 31B writing score (lmarena)
- Coding
- Behind the leadersHow often people preferred its answers to coding questions. 1,459 on Arena's scale · Source for Gemma 4 31B coding score (lmarena)
- Maths
- Not enough dataNo result yet for hard maths problems.
- Knowledge
- Not enough dataNo result yet for short factual questions.
Evidence completeness: Partial evidence, 7 of 13 tests have a result.
Where it is relatively stronger or weaker
This compares each kind of request with the model's own overall position among the 77 listed models, using Arena's votes. It is about performance on different languages, kinds of request and topics, not about political or social bias.
Relatively stronger
- Russian: 34th of 77 on questions asked in Russian, against 59th of 77 overall. Source for Russian
Relatively weaker
No category stands out as clearly weaker than its overall position.
Show every category we can see
| Category | Position | Arena score | Votes | Source |
|---|---|---|---|---|
| Expert-level prompts Kind of request | 51st of 77 | 1,465 95% range 1,440 to 1,489 | 502 | Source for Expert-level prompts |
| Hard prompts Kind of request | 62nd of 77 | 1,448 95% range 1,438 to 1,457 | 3,563 | Source for Hard prompts |
| Long prompts Kind of request | 54th of 77 | 1,446 95% range 1,432 to 1,459 | 1,796 | Source for Long prompts |
| Longer conversations Kind of request | 51st of 77 | 1,452 95% range 1,435 to 1,469 | 1,124 | Source for Longer conversations |
| Chinese Language | 63rd of 76 | 1,475 95% range 1,450 to 1,500 | 522 | Source for Chinese |
| English prompts Language | 56th of 77 | 1,451 95% range 1,439 to 1,463 | 2,329 | Source for English prompts |
| Non-English prompts Language | 60th of 77 | 1,431 95% range 1,422 to 1,440 | 3,809 | Source for Non-English prompts |
| Russian Language | 34th of 77 | 1,461 95% range 1,441 to 1,480 | 825 | Source for Russian |
| Entertainment, sport and media Topic | 63rd of 77 | 1,405 95% range 1,387 to 1,422 | 1,092 | Source for Entertainment, sport and media |
| Law and government Topic | 62nd of 76 | 1,438 95% range 1,412 to 1,464 | 452 | Source for Law and government |
| Medicine and health Topic | 51st of 75 | 1,450 95% range 1,423 to 1,478 | 422 | Source for Medicine and health |
| Science Topic | 54th of 77 | 1,457 95% range 1,438 to 1,476 | 932 | Source for Science |
| Software and IT Topic | 55th of 77 | 1,463 95% range 1,451 to 1,475 | 2,217 | Source for Software and IT |
Bias: what we can and can't see
Our sources do not measure political lean, stereotyping, flattery (agreeing with you too readily) or safety behaviour, so we cannot say how any model does on those. What we can show is above: where a model is relatively stronger or weaker across languages, kinds of request and topics. Also, people voting in Arena can favour answer style, such as length and formatting, and our numbers do not adjust for that. See what AI bias means and what these tests can tell you.
Best-supported use cases
- Homework and maths: not ranked, because there is not enough data for the main test.
- Research and checking facts: not ranked, because there is not enough data for the main test.
Where an alternative might be more suitable
- Writing and creating content: rated Behind the leaders on writing votes. Compare it with Claude Opus 5.5, the best-supported choice for this task.
- Coding and building things: rated Behind the leaders on coding votes. Compare it with Claude Opus 5.5, the best-supported choice for this task.
- Homework and maths: no result on hard maths problems, so we cannot say. See which models the evidence does support.
- Business emails and everyday work: rated Middle on business-task votes. Compare it with Gemini 4 Argon, the best-supported choice for this task.
- Research and checking facts: no result on short factual questions, so we cannot say. See which models the evidence does support.
What it is
- Maker
- Google DeepMind
- Access, as listed by Epoch AI
- Not stated
- Licence, as listed by Arena
- Apache 2.0
- Last observation in our data
- 28 Oct 2026
- Data last changed
- 29 Oct 2026, 14:36 UTC
We have not verified an official product page for this model, so we do not link to one.
How you can use it
Open weights. Arena lists its licence as Apache 2.0. In plain terms, the model's files are published under a licence that lets others run it, so it can be offered by more than one company, sometimes at lower cost. That does not guarantee that any particular app is free. Open licences differ, and some restrict commercial use, so check the licence terms before building a business on it.
How you can reach it (an app, a website, a developer API or a third-party tool) and what that costs are not covered by our benchmark data. Where we could verify prices from the maker's own pages, they are shown under Price. See models, apps and harnesses and how AI pricing works.
Price
Price not independently verified
See Gemini plans and pricing (checked 28 Oct 2026). Page opened and read.
We never infer subscription prices from API prices. New to “tokens” and “per million”? See how AI pricing works.
Good to remember
- Benchmarks are samples of tasks. None proves performance in every situation.
- Arena scores come from people choosing between answers. That measures preference, not correctness.
- Ratings are relative to the 77 models listed here. A different set would change them.
- Sources test models at different settings. Each result shows the setting used, and we do not adjust it.
The evidence, for the curious
Show all 13 tests with settings and sources
| Test | Result | Rating | Setting used | Source identifier | Observed | Links |
|---|---|---|---|---|---|---|
| Arena text: overall preference Arena 'text' leaderboard (Bradley-Terry rating, no style control) |
1,443 Arena score 95% range 1,436 to 1,450, 6,138 votes |
Behind the leaders | effort: not stated, bradley-terry rating | Gemma 4 31B lmarena · text/latest |
28 Oct 2026 | Source for Gemma 4 31B Arena text: overall preference score (lmarena) |
| Arena text: creative writing Arena 'text' leaderboard, category creative_writing |
1,416 Arena score 95% range 1,398 to 1,435, 971 votes |
Behind the leaders | effort: not stated, bradley-terry rating | Gemma 4 31B lmarena · text/latest |
28 Oct 2026 | Source for Gemma 4 31B Arena text: creative writing score (lmarena) |
| Arena text: coding Arena 'text' leaderboard, category coding |
1,459 Arena score 95% range 1,444 to 1,474, 1,411 votes |
Behind the leaders | effort: not stated, bradley-terry rating | Gemma 4 31B lmarena · text/latest |
28 Oct 2026 | Source for Gemma 4 31B Arena text: coding score (lmarena) |
| Arena text: maths Arena 'text' leaderboard, category math |
1,464 Arena score 95% range 1,438 to 1,491, 429 votes |
Middle | effort: not stated, bradley-terry rating | Gemma 4 31B lmarena · text/latest |
28 Oct 2026 | Source for Gemma 4 31B Arena text: maths score (lmarena) |
| Arena text: following instructions Arena 'text' leaderboard, category instruction_following |
1,434 Arena score 95% range 1,420 to 1,447, 1,799 votes |
Middle | effort: not stated, bradley-terry rating | Gemma 4 31B lmarena · text/latest |
28 Oct 2026 | Source for Gemma 4 31B Arena text: following instructions score (lmarena) |
| Arena text: business and management Arena 'text' leaderboard, category industry_business_and_management_and_financial_operations |
1,430 Arena score 95% range 1,414 to 1,447, 1,179 votes |
Middle | effort: not stated, bradley-terry rating | Gemma 4 31B lmarena · text/latest |
28 Oct 2026 | Source for Gemma 4 31B Arena text: business and management score (lmarena) |
| Arena WebDev: building web apps Arena 'webdev' leaderboard (Code Arena), category overall |
1,369 Arena score 95% range 1,363 to 1,375, 15,489 votes |
Behind the leaders | effort: not stated, bradley-terry rating | Gemma 4 31B lmarena · webdev/latest |
28 Oct 2026 | Source for Gemma 4 31B Arena WebDev: building web apps score (lmarena) |
| FrontierMath Tiers 1-3 (v2) | Not enough data | |||||
| OTIS Mock AIME 2024-2025 | Not enough data | |||||
| GPQA Diamond | Not enough data | |||||
| SimpleQA Verified | Not enough data | |||||
| SWE-bench Verified (Epoch runs) | Not enough data | |||||
| FrontierCode (reported by Cognition) | Not enough data | |||||
Missing evidence
Not enough data for these tests:
- FrontierMath Tiers 1-3 (v2)
- OTIS Mock AIME 2024-2025
- GPQA Diamond
- SimpleQA Verified
- SWE-bench Verified (Epoch runs)
- FrontierCode (reported by Cognition)
Missing results are not low scores and do not count against the model.
Where a source gives no public run log, we cannot show the run behind the score. See data sources.
Not sure this is the right one?
Tell us what you need help with and we'll narrow the options down.