Olmo 3.1 32B Think ranks first on 2 of the 23 benchmarks it has scores on.
Ai2 Dec 2025 Open weights Flagship
Its best score on each benchmark, and where that ranks among every model with a score there. We rank each model once, at its best setting.
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| AGIEval | 89.2% | #1 of 35 | Model card | |
| IFEval | 93.8% | #9 of 126 | Model card | |
| BIG-Bench Hard | 88.6% | #9 of 71 | Model card | |
| AlpacaEval 2.0 | 69.1% | #4 of 16 | Model card | |
| MMLU | 86.4% | #46 of 127 | Model card | |
| PopQA | 30.9% | #4 of 7 | Model card | |
| IFBench | 68.1% | #47 of 70 | Model card | |
| LMArena (Elo) | 1,286 | #242 of 333 | Official leaderboard | |
| Longform Creative Writing (EQ-Bench) | 47.0 | #89 of 119 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MATH | 96.2% | #2 of 93 | Model card | |
| AIME 2024 | 80.6% | #40 of 87 | Model card | |
| OMEGA | 53.4% | #1 of 2 | Model card | |
| AIME 2025 | 78.1% | #92 of 143 | Model card | |
| ZebraLogic | 80.1% | #17 of 23 | Model card | |
| GPQA Diamond | 57.5% | #214 of 284 | Model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| LiveCodeBench | 83.3% | #7 of 73 | Model card | |
| HumanEval+ | 91.5% | #7 of 27 | Model card | |
| MBPP+ | 68.3% | #12 of 15 | Model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| SnakeBench | 22 | #107 of 183 | Official leaderboard |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| WildGuardTest | 96.9% | #1 of 5 | Model card | |
| HarmBench | 89.7% | #1 of 4 | Model card | |
| BBQ (Bias Benchmark for QA) | 85.5% | #11 of 24 | Model card | |
| StrongREJECT | 90.5% | #11 of 22 | Model card | |
| DoAnythingNow (DAN) | 54.7% | #1 of 2 | Model card | |
| TrustLLM JailbreakTrigger | 86.4% | #1 of 2 | Model card | |
| XSTest | 91.8% | #10 of 16 | Model card | |
| Tulu 3 Safety Eval | 83.6% | #4 of 6 | Model card | |
| WMDP | 32.7% | #3 of 3 | Model card |
The benchmarks the lab chose to headline, with the scores it gave.
Every result in the longer documents the lab published with it. 10 of them aren't in the launch post.