benchmarks
Magistral Small 2506 is Mistral AI’s 24-billion-parameter, open-weight reasoning model, released under the Apache 2.0 license for multilingual, step-by-step problem solving. A newer Magistral Small 2509 generation adds image understanding; the specifications here describe the original text-only 2506 release.
Benchmark results
Every figure links to where it was published. Compare like with like — harnesses and prompt setups differ between vendors.
| Benchmark | Score | Source |
|---|---|---|
| AIMECompetition mathematics | 70.7%Jun 2025 | View |
| GPQA DiamondPhD-level science questions | 56.1%6 Aug 2026 | View |
| OTIS Mock AIME 2024-25Competition mathematics | 30%6 Aug 2026 | View |
Weighted mean of 3 verified benchmark results: AIME 70.7%, GPQA Diamond 56.1%, OTIS Mock AIME 2024-25 30%.
Capabilities
- Dedicated reasoning for multi-step mathematical and analytical problems.
- Multilingual reasoning, including the ability to switch languages during problem solving.
- Open weights under Apache 2.0 support local deployment and customization.
- Supports long text inputs within a nominal 128,000-token context window.
- Can run on a single RTX 4090 or a Mac with 32 GB of memory when quantized.
What improved
- Adds dedicated reasoning training to the Mistral Small 3.1 base model.
- Uses reinforcement learning to develop extended, step-by-step problem solving.
- Extends trained reasoning behavior across multiple languages rather than restricting it to English.
Known flaws & limitations
- The recommended context limit is 40,000 tokens, below the nominal 128,000-token window.
- Deployment on a single RTX 4090 or a Mac with 32 GB of memory requires quantization.
- Its published AIME 2024 result shows that mathematical answers remain fallible.
- The 2506 release is text-only and does not provide the image understanding introduced in Magistral Small 2509.
Risks, safety & security issues
- Not yet confirmed by public reporting.
Good fit for
- Self-hosted multilingual reasoning applications.
- Mathematical and analytical assistance with independent answer checking.
- Research and customization using openly available model weights.
- Text-based workflows that benefit from extended problem solving.
Poor fit for
- Image-based reasoning or visual document analysis with the 2506 release.
- High-stakes decisions requiring guaranteed factual or mathematical correctness.
- Workloads that require optimal performance across the entire nominal 128,000-token context.
- Unquantized deployment on memory-constrained consumer hardware.
References
Profile last reviewed 10/19/2026 and re-checked automatically every few hours against the vendor's published documentation, papers and public leaderboards. Scores come from the cited benchmark results above — confirm limits and pricing against the provider's own documentation before building on them.