Skip to main content

Model profile

Gemma 4 31B

GoogleCurrentReleased Apr 2, 2026
Data verified
Overall Score
61.08Public #43 of 200Verified #33 of 99
Arena Elo
1451
Eligible category ranks
2of 8
Price (1M tokens)
$0 in / $0 out
API pricing
Speed
Not listed
Context
256K

Evidence coverage

29 of 323 tracked benchmarks are published. 7 are verified and 22 provisional. 6 of 8 categories are measured.

Updated July 23, 2026Methodology
Published / tracked
29 / 323
Verified
7
Provisional
22
Categories with evidence
6 / 8

Evidence by category

  • Agentic9 benchmarks
    Mixed evidence
  • Coding4 benchmarks
    Mixed evidence
  • Reasoning2 benchmarks
    Reported
  • Knowledge11 benchmarks
    Mixed evidence
  • Math0 benchmarks
    Not measured
  • Multilingual0 benchmarks
    Not measured
  • Multimodal2 benchmarks
    Mixed evidence
  • Inst. Following1 benchmark
    Reported
Open WeightSelf-hostReasoning
Confidence:
Low
31b

Gemma 4 31B ranks #43 out of 200 models on the public leaderboard with an overall score of 61.08/100. It also ranks #33 out of 99 on the verified leaderboard. While not a frontier model, it offers specific advantages depending on the use case.

Gemma 4 31B is a open weight model with a 256K token context window. It uses explicit chain-of-thought reasoning, which typically improves performance on math and complex reasoning tasks at the cost of higher latency and token usage.

Gemma 4 31B sits inside the Gemma 4 family alongside Gemma 4 26B A4B, Gemma 4 12B, Gemma 4 E2B, Gemma 4 E4B. This profile currently has 29 of 323 tracked benchmarks. BenchLM only exposes non-generated benchmark rows publicly, so missing categories stay blank until a sourced evaluation is available.

Its strongest category is Coding (#71), while its weakest is Agentic (#114). This performance profile makes it particularly well-suited for software development and code generation tasks.

Peer position

Exact provisional scores and ranks for the closest listed peers. A score can appear before a model clears the evidence threshold for a rank, so equal scores can have different rank states.

Range 60.5661.31

  1. Gemini 3 Pro Deep Think
    Google
    #4161.31
    Gemini 3 Pro Deep Think is #41 with a score of 61.31.
    Compare
  2. GLM-4.7
    Z.AI
    #4261.16
    GLM-4.7 is #42 with a score of 61.16.
    Compare
  3. Gemma 4 31BCurrent model
    Google
    #4361.08
    Gemma 4 31B is #43 with a score of 61.08.
  4. GPT-5.4 Pro
    OpenAI
    #4460.89
    GPT-5.4 Pro is #44 with a score of 60.89.
    Compare
  5. Qwen3.5-27B
    Alibaba
    #4560.7
    Qwen3.5-27B is #45 with a score of 60.7.
    Compare
  6. DeepSeek V4 Pro
    DeepSeek
    #4660.66
    DeepSeek V4 Pro is #46 with a score of 60.66.
    Compare
  7. Qwen3.5-122B-A10B
    Alibaba
    #4760.56
    Qwen3.5-122B-A10B is #47 with a score of 60.56.
    Compare

Category percentile

More

Relative position among models eligible for each sourced category. A higher percentile means a stronger position within that category's ranked cohort; 100 is highest.

  1. Coding42%
    Eligible cohort rank #71 of 122Category score 49.1
  2. Agentic4%
    Eligible cohort rank #114 of 119Category score 27.0

Category evidence

Scores and ranks appear only where this model has published benchmark evidence. Categories without displayable source records remain not measured.

Category scores, ranks, weighting, benchmark coverage, and evidence status
CategoryScore
AgenticRank #114 of 119Percentile 4thWeight 22%9 benchmarksMixed sources27.0
CodingRank #71 of 122Percentile 42ndWeight 20%4 benchmarksMixed sources49.1
ReasoningRank Not rankedWeight 17%2 benchmarksReported66.4
KnowledgeRank Not rankedWeight 12%11 benchmarksMixed sources51.5
MathWeight 5%0 benchmarksNot measuredNot measured
MultilingualWeight 7%0 benchmarksNot measuredNot measured
MultimodalRank Not rankedWeight 12%2 benchmarksMixed sources59.4
Inst. FollowingWeight 5%1 benchmarkReportedScore pending

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
Model the full break-even

Chatbot Arena performance

Scroll horizontally to inspect confidence intervals and vote counts.

Chatbot Arena Elo, confidence interval, and vote count by evaluation view
ViewEloConfidence intervalVotes
Text Overall1451±7.65,879
Coding1498±15.41,362
Math1470±27.5399
Instruction Following1452±14.11,663
Creative Writing1421±19.3937
Multi-turn1464±17.81,069
Hard Prompts1473±9.93,368
Hard Prompts (English)1483±14.71,530
Longer Query1467±13.91,656

Benchmark Details

Rows below have a displayable published verification record. Each source link and provenance note remains in the page HTML while its category is closed. Source-unverified manual rows and generated rows stay hidden.

Agentic9 benchmarks
AA Agentic IndexReported

Artificial Analysis Agentic Index

14.4%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
τ²-bench resultsReported

τ²-Bench Tool-Agent-User Evaluation

59.9%Display only
Source: Artificial Analysis: tau2-bench leaderboardProvenance: Display-only row synced from the current Artificial Analysis evaluation leaderboard. It is excluded from BenchLM weighted scoring.
GDPval-AAReported

GDPval-AA normalized

15.2%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
GDPval-AAReported
804Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
Gert LabsBenchmark exact

Gert Labs Composite Game Benchmark

35.26%Display only
Source: Gert Labs rankingsProvenance: Gert Labs reports this composite leaderboard score in the public rankings API. BenchLM scales the source gscore from 0-1 to 0-100 and stores it as a display-only agentic benchmark.
AA EnterpriseOps-GymReported

Artificial Analysis EnterpriseOps-Gym

28.3%Display only
Source: Artificial Analysis: enterprise-ops-gym-aa leaderboardProvenance: Display-only row synced from the current Artificial Analysis evaluation leaderboard. It is excluded from BenchLM weighted scoring.
AA ITBenchReported

Artificial Analysis ITBench-AA

37.3%Display only
Source: Artificial Analysis: itbench-aa leaderboardProvenance: Display-only row synced from the current Artificial Analysis evaluation leaderboard. It is excluded from BenchLM weighted scoring.
AA Tau3 BankingReported

Artificial Analysis Tau3-Banking

15.1%Display only
Source: Artificial Analysis: tau3-banking leaderboardProvenance: Display-only row synced from the current Artificial Analysis evaluation leaderboard. It is excluded from BenchLM weighted scoring.
terminalBenchHardReported
36.4%Display only
Source: Artificial Analysis: terminalbench-hard leaderboardProvenance: Display-only row synced from the current Artificial Analysis evaluation leaderboard. It is excluded from BenchLM weighted scoring.
Coding4 benchmarks
SWE-RebenchBenchmark exact
41.6%Weighted 20%
Source: SWE-Rebench leaderboardProvenance: Live default SWE-Rebench leaderboard lists Gemma 4 31B at 41.6% resolved rate.
React Native EvalsBenchmark exact
75.2%Display only
Source: React Native Evals leaderboardProvenance: React Native Evals reports this exact overall score for Gemma 4 31B It in the public dashboard run finished on 2026-04-28.
AA Coding IndexReported

Artificial Analysis Coding Index

43.4%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
AA-SciCodeReported

Artificial Analysis SciCode

43.4%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
Reasoning2 benchmarks
AA-LCRReported

Artificial Analysis Long Context Reasoning

62.0%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
CritPtReported

Critical Physics Tasks

1.4%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
Knowledge11 benchmarks
HLEProvider exact

Humanity's Last Exam

26.5%Weighted 45%
Source: Gemma 4 31B model cardProvenance: Provider exact
MMLU-ProProvider exact

Massive Multitask Language Understanding Professional

85.2%Weighted 30%
Source: Gemma 4 31B model cardProvenance: Provider exact
GPQAReported

Graduate-Level Google-Proof Q&A

84.3%Weighted 7%
Source: Reported upstream sourceProvenance: Reported row carried from an upstream public source. Displayable on BenchLM, but not treated as verified unless explicitly marked otherwise.
HLE w/o toolsProvider exact

Humanity's Last Exam without tools

19.5%Display only
Source: Gemma 4 31B model cardProvenance: Provider exact
Artificial Analysis Intelligence IndexReported
29.4%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
AA-GPQA DiamondReported

Artificial Analysis GPQA Diamond

85.7%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
AA-HLEReported

Artificial Analysis Humanity's Last Exam

22.7%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
AA-Omniscience IndexReported

Artificial Analysis Omniscience Index

-45.4%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
AA-Omniscience AccuracyReported

Artificial Analysis Omniscience Accuracy

19.9%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
AA-Omniscience Hallucination RateReported

Artificial Analysis Omniscience Hallucination Rate

81.6%Display only
Source: Artificial Analysis model benchmarksProvenance: Display-only row synced from the current Artificial Analysis model payload. It is excluded from BenchLM weighted scoring.
AA Openness IndexReported

Artificial Analysis Openness Index

38.9%Display only
Source: Artificial Analysis: artificial-analysis-openness-index leaderboardProvenance: Display-only row synced from the current Artificial Analysis evaluation leaderboard. It is excluded from BenchLM weighted scoring.
Multimodal2 benchmarks
MMMU-ProProvider exact

Massive Multi-discipline Multimodal Understanding Pro

76.9%Weighted 45%
Source: Gemma 4 31B model cardProvenance: Provider exact
AA-MMMU-ProReported

Artificial Analysis MMMU-Pro

73.4%Display only
Source: Artificial Analysis: mmmu-pro leaderboardProvenance: Display-only row synced from the current Artificial Analysis evaluation leaderboard. It is excluded from BenchLM weighted scoring.
Inst. Following1 benchmark
AA-IFBenchReported

Artificial Analysis IFBench

75.6%Display only
Source: Artificial Analysis: ifbench leaderboardProvenance: Display-only row synced from the current Artificial Analysis evaluation leaderboard. It is excluded from BenchLM weighted scoring.

Frequently Asked Questions

How does Gemma 4 31B perform overall in AI benchmarks?

Gemma 4 31B has 29 published benchmark scores on BenchLM, but it does not yet have enough non-generated coverage to receive a global overall rank.

Is Gemma 4 31B good for knowledge and understanding?

Gemma 4 31B has visible benchmark coverage in knowledge and understanding, but BenchLM does not currently assign it a global category rank there.

Is Gemma 4 31B good for coding and programming?

Gemma 4 31B ranks #71 out of 122 models in coding and programming benchmarks with an average score of 49.1. There are stronger options in this category.

Is Gemma 4 31B good for reasoning and logic?

Gemma 4 31B has visible benchmark coverage in reasoning and logic, but BenchLM does not currently assign it a global category rank there.

Is Gemma 4 31B good for agentic tool use and computer tasks?

Gemma 4 31B ranks #114 out of 119 models in agentic tool use and computer tasks benchmarks with an average score of 27. There are stronger options in this category.

Is Gemma 4 31B good for multimodal and grounded tasks?

Gemma 4 31B has visible benchmark coverage in multimodal and grounded tasks, but BenchLM does not currently assign it a global category rank there.

Is Gemma 4 31B good for instruction following?

Gemma 4 31B has visible benchmark coverage in instruction following, but BenchLM does not currently assign it a global category rank there.

Is Gemma 4 31B open source?

Yes, Gemma 4 31B is an open weight model created by Google, meaning it can be downloaded and run locally or fine-tuned for specific use cases.

Which sibling models are related to Gemma 4 31B?

Gemma 4 31B belongs to the Gemma 4 family. Related variants on BenchLM include Gemma 4 26B A4B, Gemma 4 12B, Gemma 4 E2B, Gemma 4 E4B.

Does Gemma 4 31B have full benchmark coverage on BenchLM?

Not yet. Gemma 4 31B currently has 29 published benchmark scores out of the 323 benchmarks BenchLM tracks. BenchLM only exposes non-generated public benchmark rows, so missing categories stay blank until a sourced evaluation is available.

What is the context window size of Gemma 4 31B?

Gemma 4 31B has a published context window of 256K, which determines how much text it can process in a single interaction.

Last updated: July 23, 2026 · Runtime metrics stay blank until BenchLM has a sourced snapshot.

Choose with this week’s evidence

Join 2,000+ readers for ranking moves, new releases, pricing changes, and the evidence behind them.

Free. One email per week.