Mistral AI Announces Magistral, Its First Reasoning Model, in Open and Enterprise Variants
Mistral AI
- Models
- Open source
- Benchmarks
- #magistral
- #mistral
- #reasoning
- #aime2024
Mistral AI announced Magistral, described as its first reasoning model, released as a 24B-parameter open-weight Small version and an enterprise Medium version. The company reports AIME2024 scores of 70.7% (Small) and 73.6% (Medium), rising to 83.3% and 90% with majority voting @64, and emphasizes multilingual chain-of-thought, traceable reasoning, and Le Chat speed features.
What Mistral announced
Mistral AI's blog post introduces Magistral as the company's first reasoning model, released in two variants. Magistral Small is a 24-billion-parameter version offered as open weights under Apache 2.0 for self-deployment via Hugging Face. Magistral Medium is an enterprise version available as a preview in Le Chat or through the La Plateforme API, and Mistral says it is also on Amazon SageMaker, with IBM WatsonX, Azure AI and Google Cloud Marketplace to follow. The company reports that Medium scored 73.6% on AIME2024 and 90% with majority voting @64, while Small scored 70.7% and 83.3% respectively. Mistral frames the model around multi-step logic, reasoning traces that users can follow in their own language, and support for languages including English, French, Spanish, German, Italian, Arabic, Russian and Simplified Chinese. It also cites Le Chat's Think mode and Flash Answers, claiming up to 10x faster token throughput than most competitors. These are the publisher's own claims from a short excerpt, not independent evaluations.
A scenario for a compliance analyst
Consider a compliance analyst at a financial firm reviewing rule-based decisions, such as whether a set of transactions satisfies a policy. Because Mistral positions Magistral for traceable reasoning in regulated sectors, a team could trial the model on a small set of already-reviewed cases, asking it to walk through the rule logic and language of the source documents. The analyst would compare the model's reasoning trace against the documented policy steps, rather than treating the final answer as authoritative. This uses the announced capability as intended without assuming any accuracy or compliance outcome; the excerpt provides no evidence of performance on such tasks.
Trial steps and success checks
A team could select ten to twenty past cases with known correct outcomes, run them through Magistral Medium in the preview environment (or Small if self-hosting is required), and record the reasoning trace, final answer, and language used. Set a baseline first using an existing model or manual review, then compare on agreement with the known outcome and on how easily an analyst can follow and verify each step. Failure checks should include cases where the trace looks plausible but contradicts the final answer, and cases in a non-English language to test the multilingual claim. A human reviewer should sign off on any case before it influences a real decision, since the post offers no independent accuracy data.
Trade-offs and when to hold back
Self-deploying Magistral Small means running a 24B-parameter model, which carries infrastructure and oversight costs; the Medium version is a preview, and Mistral itself says it aims to iterate quickly, so behavior may shift. The excerpt lists no pricing, no independent benchmarks, and no detail on data handling, so teams with strict procurement or privacy rules should wait for more information. Majority-voting results reported by Mistral require multiple runs, which multiplies cost and latency. If a use case needs only simple lookups or creative output without verifiable steps, a reasoning model's longer deliberation may not be worth the overhead. Treat the speed comparison as a vendor claim and test it in the target environment.
Context and limitations
- Based on a short source excerpt; the full article and underlying paper were not analyzed.
- Benchmark figures (AIME2024, majority voting, speed claims) are Mistral's own reported numbers, not independently verified.
- Preview and 'soon available' statements reflect the announcement date and may change.
- Based on bounded source text, not independent testing. Practical suggestions are editorial proposals.