Google DeepMind·released 2026-04-021 source
Gemma 4 31B IT benchmark scores: 8 benchmarks tracked. 38% of its scores are independently reproduced — the rest are self-reported or unverified.
Leads on 0 of 8 benchmarks·best result 73.31 on OTIS Mock AIME 2024-2025 (reproduced)·3 of 8 independently reproduced
| Benchmark | Score | Evidence status | Measured |
|---|---|---|---|
| OTIS Mock AIME 2024-2025 | 73.31accuracy (%) | reproduced· optimizedT1 | 2026-04-02Epoch AI |
| DTBench | 71.12accuracy (%) | unverifiedT1 | 2026-04-02https://conceptualreasoning.ai/dtbench |
| GPQA diamond | 67.68accuracy (%) | reproduced· optimizedT1 | 2026-04-02Epoch AI |
| WeirdML | 52.26accuracy (%) | unverifiedT1 | 2026-04-02https://htihle.github.io/weirdml.html |
| LMCA | 46.18accuracy (%) | unverifiedT1 | 2026-04-02https://conceptualreasoning.ai/lmca |
| Surface Evolver Bench | 30.62accuracy (%) | unverifiedT1 | 2026-04-02https://yhenon.github.io/surface-evolver-llm-eval/ |
| SimpleQA Verified | 10.4accuracy (%) | unverifiedT2 | 2026-04-02 |
| Chess Puzzles | 0.04accuracy (%) | reproducedT1 | 2026-04-02Epoch AI |
Claims drawn from cited facts, not live model generation.
Developed by Google DeepMind and released in 2026, this model is a mid-size system with 31 billion parameters. At this model's scale, the trade-off is clear: smaller models offer more efficiency and lower cost, while larger models provide more headroom at greater expense.
4 cited facts
This model is scored on 8 benchmarks, and it currently holds no top score on any of them. Its record can be taken at face value, because the scores have been independently reproduced: outside evaluation backs the numbers up rather than resting on vendor claims alone.
3 cited facts
Averaging a 43.57-point deficit to the top score across its benchmarks, this model sits well behind the leaders rather than at the frontier. It also holds the top score on none of those benchmarks, meaning it has no category leadership yet.
2 cited facts
2 facts cross-checked across data sources: 2 single-source.
Source facts, citations, and refresh stamp for this record.
Sources: curated_param_counts, epoch_benchmarks
Gemma 4 31B IT is an AI model developed by Google DeepMind, released 2026-04-02. Its benchmark record below tags every score as independently reproduced, self-reported, or unverified.
Gemma 4 31B IT has recorded scores on 8 benchmarks, each shown with its evidence status.
Gemma 4 31B IT has recorded scores on 8 benchmarks — GPQA diamond, OTIS Mock AIME 2024-2025, DTBench, WeirdML, LMCA, SimpleQA Verified, and 2 more. The full table above shows each score with its evidence status.
3 of 8 recorded scores (38%) are independently reproduced rather than self-reported by the lab.
Gemma 4 31B IT has 8 tracked claims: 3 independently reproduced, 5 unverified.