Devstral Small 2: benchmark results
Mistral's Apache-2.0 24B agentic coding model with a 256K context, strong on SWE-bench Verified for its size (December 2025). Provider: Mistral. Released 2025-12-09. Access: Open.
Unified ELO 1558 ± 1, rank #408 of 1392 rated models, from 103 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Enkokilish Bench (English) | 84.85 | Riddles Solved (%) | 100 |
| Konkur 1404 - Art | 55.17 | Accuracy (%, text-only) | 94.7 |
| Konkur 1404 - Humanities | 45.7 | Accuracy (%, text-only) | 89.5 |
| CringeBench | 0.74 | Cringe Score (0-10, lower is better) | 84.8 |
| Konkur 1404 - Overall | 43.8 | Accuracy (%, text-only) | 84.2 |
| UGI Leaderboard | 45.28 | UGI Score | 81 |
| Konkur 1404 - Mathematics | 44.35 | Accuracy (%, text-only) | 76.3 |
| Konkur 1404 - Experimental Sciences | 39.13 | Accuracy (%, text-only) | 68.4 |
| UGI - Willingness (W/10) | 6.8 | W/10 Score | 67 |
| Konkur 1404 - Foreign Language | 37.5 | Accuracy (%, text-only) | 65.8 |
| AA Omniscience - Software Engineering (SWE) - Kotlin | 20 | Accuracy (%) | 63 |
| AA Omniscience - Software Engineering (SWE) - HTML | 32 | Accuracy (%) | 60.4 |
Interactive version: theaggregate.ai/model?slug=devstral-small-2 · How It Works · Data refreshed daily, snapshot 2026-09-05.