Overview
Magistral 1.2 is Mistral AI's reasoning line, shipped on 17 September 2025 as a pair of models: the open-weight Magistral Small 1.2 (a 24-billion-parameter model under Apache 2.0, distributed as mistralai/Magistral-Small-2509) and the closed, API-only Magistral Medium 1.2 (magistral-medium-2509). Both are 'thinking' models that work through problems with an explicit chain-of-thought trace before answering, positioning them as Mistral's answer to OpenAI's o-series and DeepSeek-R1.
The headline change in 1.2 is multimodality: both models now carry a vision encoder, so they reason over images as well as text — useful for chart reading, diagram interpretation, document layout analysis, and visual question answering. Mistral also reports roughly 15% gains on math and coding benchmarks (AIME 24/25 and LiveCodeBench) over the 1.1 release, plus better tool use (web search, code interpreter, image generation) and clearer, better-formatted responses. Both share a 128K-token context window and span 24 languages.
Magistral Small 1.2 is small enough to run locally — Mistral positions it for a single RTX 4090 GPU or a 32GB-RAM MacBook once quantized — while Magistral Medium 1.2 is the frontier-class variant that powers reasoning in Le Chat and is served through Mistral's La Plateforme API. Magistral Medium 1.2 is the more capable of the two on every published benchmark.
| Released | 2025-09-17 |
|---|---|
| License | Apache 2.0 (Magistral Small 1.2); proprietary (Magistral Medium 1.2) |
| Weights | Open weights |
| Parameters | 24B (Magistral Small 1.2); undisclosed (Magistral Medium 1.2) |
| Context | 128K |
| Max output | 128K (shared context budget; no separate output cap published) |
| Architecture | Dense transformer reasoning model with chain-of-thought ([THINK]/[/THINK] traces). Magistral Small 1.2 is built on Mistral Small 3.2 (2506) with a vision encoder added, then specialized for reasoning via supervised fine-tuning and reinforcement learning. |
| Modalities | Text, Vision |
| Status | Available |
Benchmarks
- AIME 2024 (pass@1) — Magistral Medium 1.291.82%
- AIME 2025 (pass@1) — Magistral Medium 1.283.48%
- GPQA Diamond — Magistral Medium 1.276.26%
- LiveCodeBench v5 — Magistral Medium 1.275%
- AIME 2024 (pass@1) — Magistral Small 1.286.14%
- AIME 2025 (pass@1) — Magistral Small 1.277.34%
- GPQA Diamond — Magistral Small 1.270.07%
- LiveCodeBench v5 — Magistral Small 1.270.88%
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Pricing
| Input | $2.00 / 1M tokens (Magistral Medium); $0.50 / 1M tokens (Magistral Small) per 1M tokens |
|---|---|
| Output | $5.00 / 1M tokens (Magistral Medium); $1.50 / 1M tokens (Magistral Small) per 1M tokens |
La Plateforme API pricing. Magistral Small is also free to self-host under Apache 2.0.
Strengths
- Transparent chain-of-thought reasoning with explicit [THINK] traces, strong on math (AIME) and competitive coding (LiveCodeBench)
- Open-weight 24B Small variant under Apache 2.0 — free to self-host, fine-tune, and run on a single consumer GPU or a 32GB MacBook
- New vision encoder in 1.2 lets both models reason over images, charts, and document layouts
- 128K-token context and broad multilingual coverage (24 languages)
- Cheaper API pricing than most frontier reasoning models ($0.50/$1.50 per 1M tokens for Small; $2/$5 for Medium)
Best for
- Multi-step math, logic, and STEM problem solving where a visible reasoning trace is valuable
- Competitive-programming-style and algorithmic coding tasks
- Multimodal reasoning over diagrams, charts, screenshots, and document layouts
- On-device / private reasoning assistants using the open-weight 24B Small model
- Multilingual reasoning and agentic workflows with web search and code-interpreter tools
How to access
| Provider | Model ID |
|---|---|
| Mistral AI (La Plateforme) ↗ | magistral-medium-2509 |
| Mistral AI (La Plateforme) ↗ | magistral-small-2509 |
Magistral — every version
The full lineage of the Magistral line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.
| Version | Released | Context | License |
|---|---|---|---|
| Magistral 1.2 (Small & Medium)current | 2025-09-17 | — | Open weights |
| Magistral 1.1 (Small & Medium) | 2025-07 | — | Open weights |
| Magistral 1.0 (Small & Medium) | 2025-06-10 | — | Apache-2.0 |
FAQ
What's new in Magistral 1.2 versus 1.1?
The biggest change is multimodality: both Magistral Small 1.2 and Medium 1.2 add a vision encoder, so they can reason over images as well as text. Mistral also reports roughly 15% gains on math (AIME 24/25) and coding (LiveCodeBench) benchmarks, better tool use with web search and a code interpreter, and clearer, better-formatted responses.
Is Magistral 1.2 open source?
Partly. Magistral Small 1.2 (24B parameters) is open-weight under the Apache 2.0 license and downloadable from Hugging Face as mistralai/Magistral-Small-2509, so you can self-host and fine-tune it commercially. Magistral Medium 1.2 is proprietary and available only through Mistral's La Plateforme API and Le Chat.
Can Magistral Small 1.2 run locally?
Yes. At 24B parameters it is small enough to run on a single Nvidia RTX 4090 GPU, and Mistral says a quantized version fits in a MacBook with 32GB of RAM, making it one of the few reasoning models you can run fully on-device.
How much does the Magistral 1.2 API cost?
On Mistral's La Plateforme, Magistral Medium 1.2 costs $2.00 per million input tokens and $5.00 per million output tokens, while Magistral Small 1.2 costs $0.50 input and $1.50 output per million tokens. Magistral Small is also free to run yourself under Apache 2.0.