Gemma 4 released
Google · Model release
Source confirmed
Data as of September 30, 2026 · How the score is built
EstablishedOpen WeightReasoning
Gemma 4 31B
Decision readingGemma 4 31B scores 44.8 out of 100 and ranks #107 of 210. This profile shows 8 source-displayable benchmark rows; its strongest eligible category is Multimodal & Grounded at #32. Published weights can be self-hosted, but infrastructure cost varies and is not a comparable API token rate.
Released Apr 2, 2026 — see all recent releases
Gemma 4 31B will be repriced, updated or retired. Get each notice with its source and date. Follow model changes
Share or export
Decision snapshot
Each value carries a field reference instead of floating alone. Markers compare this model with the current ranked and priced catalog; they are not absolute quality thresholds.
Capability
44.8/100
field median 50.6#107 of 210 ranked models
Public
#107of 210
Verified —
Price
Self-hosted; infrastructure cost varies
input median $1No comparable first-party hosted token rate
Speed
36tok/s
field median 92 tok/sFirst token 49.83 s
Context
256Ktokens
field median 256,000Maximum output length is tracked separately
Strongest published evidence
Multimodal & Grounded ranks #32. Particularly strong for screenshots, documents, charts, and grounded multimodal workflows.
Validate before choosing
8 published rows leave some tracked benchmark slots empty. No comparable first-party API token rate is published.
Source-linked · 8 displayable benchmark rows
Follow model changesCategory score record
Scores and ranks appear only where published evidence can be displayed. The table keeps the score, weight, cohort, and evidence state together.
| Category | Score | Rank | Percentile | Weight | Benchmarks | Evidence |
|---|---|---|---|---|---|---|
| AgenticRank #92 of 117Percentile 22ndWeight 22%1 benchmarkVerified | 24.5 | 1 benchmark | Verified | |||
| CodingRank #80 of 143Percentile 44thWeight 20%2 benchmarksVerified | 35.6 | 2 benchmarks | Verified | |||
| ReasoningWeight 17%0 benchmarksNot measured | Not measured | 0 benchmarks | Not measured | |||
| MultimodalRank #32 of 50Percentile 37thWeight 12%1 benchmarkVerified | 59.5 | 1 benchmark | Verified | |||
| KnowledgeRank #99 of 169Percentile 42ndWeight 12%4 benchmarksMixed sources | 40.0 | 4 benchmarks | Mixed sources | |||
| MultilingualWeight 7%0 benchmarksNot measured | Not measured | 0 benchmarks | Not measured | |||
| Inst. FollowingWeight 5%0 benchmarksNot measured | Not measured | 0 benchmarks | Not measured | |||
| MathWeight 5%0 benchmarksNot measured | Not measured | 0 benchmarks | Not measured |
8 of 491 tracked benchmark slots have displayable evidence · bars run 0–100
Coverage detailsHow much of this is verified
Coverage is split by category so a strong number never hides a thin evidence base. Verified means the row is tied to a published source; provisional rows remain visible but separate.
- Agentic1/1 verified
- Coding2/2 verified
- ReasoningNot measured
- Multimodal1/1 verified
- Knowledge3/4 verified
- MultilingualNot measured
- Inst. FollowingNot measured
- MathNot measured
Verified sourceProvisionalNot measured
Capability shape
Each axis shows percentile within that category’s eligible cohort. The comparison outline is the median of the six nearest public-score peers; a collapsed vertex means the category is not rank-eligible.
Gemma 4 31B category percentile values
- Agentic22nd percentile
- Coding44th percentile
- ReasoningNot eligible
- Multimodal37th percentile
- Knowledge42nd percentile
- MultilingualNot eligible
- Instruction followingNot eligible
- MathNot eligible
The dashed outline is median of 6 nearest peers.
Eligible category ranks
- Agentic#92/117
- Coding#80/143
- ReasoningNot ranked
- Multimodal#32/50
- Knowledge#99/169
- MultilingualNot ranked
- Inst. FollowingNot ranked
- MathNot ranked
Top decileTop quartileMid-fieldNot eligible
Benchmark ledger
Coding opens by default. The marker compares each value with the best source-verified result in the catalog; provisional leaders do not set the reference. Expand the remaining categories for every published row.
Coding2 rows
| Benchmark | Score | Versus best verified row | Gap | Weight | Evidence |
|---|---|---|---|---|---|
| SWE-Rebench | Score41.6% | Versus best verified row Best verified: Claude Opus 4.6 · 65.3% | Gap23.7 behind | Weight5% ref. weight | Benchmark exact |
| React Native Evals | Score75.2% | Versus best verified row Best verified: Composer 2 · 96.1% | Gap20.9 behind | WeightDisplay only | Benchmark exact |
Agentic1 row
| Benchmark | Score | Versus best verified row | Gap | Weight | Evidence |
|---|---|---|---|---|---|
| Gert LabsGert Labs Composite Game Benchmark | Score35.26% | Versus best verified row Best verified: Claude Opus 4.8 · 72.97% | Gap37.7 behind | WeightDisplay only | Benchmark exact |
Multimodal1 row
| Benchmark | Score | Versus best verified row | Gap | Weight | Evidence |
|---|---|---|---|---|---|
| MMMU-ProMassive Multi-discipline Multimodal Understanding Pro | Score76.9% | Versus best verified row Best verified: Gemini 3.5 Flash · 83.6% | Gap6.7 behind | WeightWeighted 40% | Provider exact |
Knowledge4 rows
| Benchmark | Score | Versus best verified row | Gap | Weight | Evidence |
|---|---|---|---|---|---|
| HLEHumanity's Last Exam | Score26.5% | Versus best verified row Best verified: Claude Fable 5.1 · 65% | Gap38.5 behind | Weight44% ref. weight | Provider exact |
| HLE w/o toolsHumanity's Last Exam without tools | Score19.5% | Versus best verified row Best verified: Claude Opus 5.5 · 64.4% | Gap44.9 behind | Weight7% ref. weight | Provider exact |
| MMLU-ProMassive Multitask Language Understanding Professional | Score85.2% | Versus best verified row Best verified: Qwen3.7 Max · 89.6% | Gap4.4 behind | Weight6% ref. weight | Provider exact |
| GPQAGraduate-Level Google-Proof Q&A | Score84.3% | Versus best verified row Best verified: GPT-6 Astra · 96% | Gap11.7 behind | Weight3% ref. weight | Reported |
Bars run 0–100; the dark tick marks the best source-verified value
All 8 rowsLineage
The sequence follows explicit supersedes links. Each score is estimated for that model; a relative can inform a sparse estimate but never sets a floor, so a newer release can score below an earlier one. Scores and prices remain blank when the corresponding public row or first-party rate is unavailable.
Radar
Full release historyGemma 4 31B release history
Radar confirmed these at the source. Use Gemma 4 31B in your work? Explore Radar to follow supported changes and choose your alerts.
Radar
Spec sheet
Each documented value carries its source. Missing fields stay visible as not sourced or not published, rather than disappearing from the page.
- API model ID
- gemma-4-31b-itGoogle Gemma Gemini API guide
- Context window
- 256KGoogle Gemma 4 model documentation
- Maximum output
- Not sourced yet
- Knowledge cutoff
- Not sourced yet
- Input modalities
- text, imageGoogle Gemma 4 model documentation
- Output modalities
- textGoogle Gemma 4 model documentation
- Parameters
- Not sourced yet
- Availability
- Gemini API · Google AI Studio · open weightsGoogle Gemma Gemini API guide
- Cloud regions
- Not tracked yet
- Lifecycle
- activeGoogle gemma-4-31B model card
- API capabilities
- Tool calling, structured outputs, and batch support are not tracked yet
- Prompt caching
- Not documented in the pricing record
- Self-host
- Gemma 4 31B needs ~24GB VRAM (1× NVIDIA RTX 4090 (24GB)).
- Rate limits
- Not tracked yet
How to read this profile
The visual layer above carries the decisions. These notes preserve the model, ranking, coverage, and family context behind the numbers.
Gemma 4 31B ranks #107 of 210 on the public leaderboard with a score of 44.75/100. It does not yet have enough sourced coverage for a verified position.
Gemma 4 31B is a open weight model with a 256K context window. It uses an explicit reasoning mode, which can improve complex problem solving while adding latency and token use.
Official exact-value snapshot from Google's April 2, 2026 Gemma 4 model card and launch post. BenchLM maps only schema-equivalent results from the published table: MMLU Pro, GPQA Diamond, LiveCodeBench v6, HLE no tools / with search, BigBench Extra Hard, MMMU Pro, and MRCR v2. AIME 2026, Codeforces, MMMLU, OmniDocBench 1.5, MATH-Vision, MedXPertQA MM, audio benchmarks, and Tau2 (average over 3) remain out-of-schema or non-equivalent to current weighted keys.
Gemma 4 31B sits in the Gemma 4 family with Gemma 4 26B A4B, Gemma 4 12B, Gemma 4 E4B, Gemma 4 E2B. 8 of 491 tracked benchmark slots currently have displayable evidence. Missing categories stay blank.
Its strongest eligible category is Multimodal & Grounded at #32, while its lowest eligible position is Knowledge at #99. particularly strong for screenshots, documents, charts, and grounded multimodal workflows.
Last updated September 30, 2026. Runtime fields remain blank until a sourced snapshot exists.
Deployment options
Self-host and provider-specific paths stay separate from benchmark evidence so operating constraints are visible before a score becomes the whole decision.
Self-host vs API cost
Estimates at 50,000 req/day · 1000 tokens/req average.
Gemma 4 31B
API / mo$0
Self-host / mo$429
Break-even—
Questions
How does Gemma 4 31B perform overall in AI benchmarks?
Gemma 4 31B ranks #107 out of 210 models on the public BenchAlign leaderboard, with a score of 44.75/100. Its evidence status is Estimated, and this profile shows 8 source-displayable benchmark rows. The label describes evidence depth, not a provider quality claim; inspect category rows before choosing a workload.
Is Gemma 4 31B good for knowledge and understanding?
Gemma 4 31B ranks #99 out of 169 eligible models for knowledge and understanding, with a public category score of 40/100. Higher-ranked alternatives are available for workloads where this category decides the choice. Check the underlying rows before treating the aggregate as a workload guarantee.
Is Gemma 4 31B good for coding and programming?
Gemma 4 31B ranks #80 out of 143 eligible models for coding and programming, with a public category score of 35.6/100. Higher-ranked alternatives are available for workloads where this category decides the choice. Check the underlying rows before treating the aggregate as a workload guarantee.
Is Gemma 4 31B good for agentic tool use and computer tasks?
Gemma 4 31B ranks #92 out of 117 eligible models for agentic tool use and computer tasks, with a public category score of 24.5/100. Higher-ranked alternatives are available for workloads where this category decides the choice. Check the underlying rows before treating the aggregate as a workload guarantee.
Is Gemma 4 31B good for multimodal and grounded tasks?
Gemma 4 31B ranks #32 out of 50 eligible models for multimodal and grounded tasks, with a public category score of 59.5/100. Higher-ranked alternatives are available for workloads where this category decides the choice. Check the underlying rows before treating the aggregate as a workload guarantee.
Is Gemma 4 31B open source?
Gemma 4 31B is an open-weight model from Google. Its weights can be downloaded for local or hosted deployment, subject to the published license. Open weight does not automatically mean open source: training data and training code may remain private, and commercial restrictions can still apply.
Which sibling models are related to Gemma 4 31B?
Gemma 4 31B belongs to the Gemma 4 family. Related tracked variants include Gemma 4 26B A4B, Gemma 4 12B, Gemma 4 E4B, plus 1 more. A sibling link indicates shared lineage or a documented configuration relationship; it does not mean the variants have identical pricing, context limits, benchmark evidence, or deployment behavior. Compare before switching.
Does Gemma 4 31B have full benchmark coverage on BenchLM?
No. Gemma 4 31B currently has 24 source-displayable rows across 491 tracked benchmark slots. The profile exposes published, non-generated evidence and leaves missing categories blank until an exact evaluation is available. Coverage describes how much was measured; it is not a penalty added to an individual benchmark result.
What is the context window size of Gemma 4 31B?
Gemma 4 31B has a documented context window of 256K. That figure is the maximum combined prompt and retained-conversation space reported for this exact model; it is not the maximum output length. The profile keeps output limits separate because providers often publish those limits independently.