Devstral 2: benchmark results
Mistral's open-weight Devstral 2 (December 2025), a 123-billion-parameter agentic coding model in FP8 for exploring codebases, editing multiple files and powering software-engineering agents, with a 256K context. Provider: Mistral. Released 2025-12-09. Access: Open.
Aggregate Rating 1608 ± 13, rank #411 of 1656 rated models, from 133 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Agent-Diff | 86 | Assertion-weighted score (%) on Agent-Diff's 224 enterprise | 87.5 |
| Agent-Diff - Pass Rate | 74 | Pass rate (%) on Agent-Diff's 224 enterprise API tasks (Box, | 87.5 |
| AA Omniscience - Software Engineering (SWE) - Kotlin | 28 | Accuracy (%) | 80.5 |
| AA Omniscience - Software Engineering (SWE) - HTML | 40 | Accuracy (%) | 75.9 |
| AA Omniscience - Software Engineering (SWE) - PHP | 32 | Accuracy (%) | 75.9 |
| AA Omniscience - Software Engineering (SWE) - Go | 24 | Accuracy (%) | 74.1 |
| AA Omniscience - Software Engineering (SWE) - Java | 20 | Accuracy (%) | 74.1 |
| AA Omniscience - Software Engineering (SWE) - C | 44 | Accuracy (%) | 73.4 |
| AA Omniscience - Software Engineering (SWE) - Dart | 24 | Accuracy (%) | 72.2 |
| AA Omniscience - Software Engineering (SWE) - JavaScript | 33.64 | Accuracy (%) | 69.1 |
| AA Omniscience - Software Engineering (SWE) - Rust | 54 | Accuracy (%) | 68.9 |
| AA Omniscience - Software Engineering (SWE) - Julia | 16 | Accuracy (%) | 68.8 |
Interactive version: theaggregate.ai/model?slug=devstral-2 · How It Works · Data refreshed daily, snapshot 2026-10-20.