Google · AI model
Gemma 4 31B
Largest model of Google's open-weights Gemma 4 family (30.7B-parameter dense model, Apache 2.0). Text and image input, video as frames, text output, context window of up to 256K tokens. Offered free of charge on the Gemini API free tier.
Last verified Available 6 sources
Key specs
- Provider
- Family
- Gemma 4
- API identifier
- gemma-4-31b-it
- Availability
- Available
- Release date
- 2026-04-02
- License
- Open source (Apache 2.0)
- Knowledge cutoff
- January 2025
- API access
- Yes
- Consumer app
- Unknown
Capabilities
Inputs
- Text input Yes
- Image input Yes
- Audio input No Model card lists 'No Audio' for 31B; audio input only on E2B, E4B and 12B
- Video input Partial Video processed as frames; maximum 60 seconds at one frame per second
- File input Unknown
Outputs
- Text output Yes
- Image output No Model card lists text as the only output
- Audio output No Model card lists text as the only output
Features
- Reasoning Optional On the Gemini API, thinking_level high enables and minimal disables thinking; in the open weights it is toggled with the <|think|> token in the system prompt
- Tool calling Yes Function calling documented for Gemma 4 on the Gemini API; model card cites native function calling
- Web search Yes Google Search grounding documented on the Gemma-on-Gemini-API page; the pricing page lists grounding with Google Search as 'Not available' for Gemma 4
"Unknown" means we have not verified this capability yet; it does not mean the model lacks it.
Pricing
No current prices listed.
Benchmarks
| Benchmark | Setting | Score | Reported by | Measured | Verified | Source |
|---|---|---|---|---|---|---|
| AutomationBench % tasks completed correctly, higher is better | v1.0.6; max effort | 1.7 % | Third party | — | 2026-10-25 | AutomationBench AI benchmark leaderboard (Zapier) |
| GDPval-AA v2.1 Elo, higher is better | v2.1; non-reasoning; Artificial Analysis Stirrup agent harness | 547 Elo | Third party | 2026-10-25 | 2026-10-25 | GDPval-AA v2.1 Leaderboard (Artificial Analysis) |
| v2.1; reasoning; Artificial Analysis Stirrup agent harness | 619 Elo | Third party | 2026-10-25 | 2026-10-25 | GDPval-AA v2.1 Leaderboard (Artificial Analysis) | |
| Humanity's Last Exam accuracy %, higher is better | no tools; thinking mode not stated | 19.5 % | Self-reported | — | 2026-10-25 | Model card - Gemma Documentation |
| with search; thinking mode not stated | 26.5 % | Self-reported | — | 2026-10-25 | Model card - Gemma Documentation | |
| Text Arena Arena score, higher is better | Overall, style control on; effort not in model name | 1,452 | Third party | 2026-10-23 | 2026-10-25 | Text Arena leaderboard (Arena) |
Leaderboard versions (for example 1.0.5 and 1.0.6) are not comparable. Some results use fallback routing for refused steps: for "Claude Fable 5.1 (with Opus 5 Fallback)", Opus 5 handled 260 of 657 tasks. Artificial Analysis runs a separate leaderboard, AutomationBench-AA, which uses a guardrail-adjusted metric; its scores are not comparable with Zapier's. The partial_credit metric is for diagnostics only. The leaderboard shows no per-row evaluation dates.
Elo is a relative rating: values change as models are added and the scale is re-anchored, so always store the date a rating was read. For example, Anthropic's launch page gives 1846 for Claude Opus 5.5, while the leaderboard on 2026-10-12 shows 1866 (Max, Default Fallback). Earlier versions (Google cites "GDPVal-AA v2") are not comparable. Judging is done by an LLM, not by human experts. Anthropic models are listed with "Default Fallback", meaning refused steps are routed to a fallback model. AA-Briefcase and GDP.pdf are separate evaluations.
Results differ a lot by setting: with tools versus without tools (Anthropic reports 45.9% without tools and 57.4% with tools for Claude Haiku 5.5), and the full set versus the text-only subset. Providers use different judge models and harnesses (Scale uses o3-mini; Alibaba states GPT-4o). Derived sets such as HLE-Verified (a third-party revision with 1,811 scored items), HLE-Rolling and HLE-Diamond are different item sets and are not comparable. The questions are public, so contamination is possible, and community reviews have reported flawed items.
Scores measure voter preference, which can be influenced by answer style and length. The leaderboard offers a style-control adjustment, so record which view was used. Scores shift with every update, so a value is only meaningful together with its leaderboard date. Entries marked "Preliminary" or with few votes have wide confidence intervals. The same model can appear several times with different reasoning-effort suffixes (for example -high or -max). Category views (Coding, Math, Creative Writing and others) are different rankings from Overall.
One row per result, without rank numbers: each result was measured under its own setting (effort level, harness, tools, test version), as stated by the source, and scores compare only within the same setting. Self-reported results come from the provider and have not been reproduced independently.
Rankings
- #12 of 14 in Best AI for research as of 2026-10-26
- #38 of 39 in Best AI for business as of 2026-10-26
Derived from our published ranking snapshots; each ranking explains its criteria and weights.
Compare Gemma 4 31B
Guides
About Gemma 4 31B
Gemma 4 31B (gemma-4-31b-it on the Gemini API) is the largest of the five Gemma 4 sizes (E2B, E4B, 12B, 26B A4B and 31B) and has 30.7B parameters in a dense architecture. It accepts text and image input, processes video as frames (up to 60 seconds at one frame per second) and returns text, with a context window of up to 256K tokens. The weights are released under the Apache 2.0 licence; on the Gemini API, released on 2 April 2026, it is offered free of charge on the free tier only, with no paid tier. Thinking can be switched on or off.
Strengths and weaknesses
Strengths
- Open weights under the Apache 2.0 licence
- Context window of up to 256K tokens
- Native function calling and a configurable thinking mode
- Free of charge on the Gemini API free tier; Google Search grounding is documented for Gemma on the Gemini API
Weaknesses
- No audio input (audio is limited to the E2B, E4B and 12B sizes)
- On the Gemini API only a free tier is offered (no paid tier), and free-tier data is used to improve Google's products
- Smaller context window than the Gemini models in this catalog (up to 256K vs 1,048,576 tokens)
Sources
Last verified Page updated 2026-10-26