At a glance: Magistral Medium, released June 10, 2025. Proprietary, API-only. Mistral’s first reasoning model. AIME 2024: 73.6% pass@1. GPQA Diamond: 70.83%. $2.00/$5.00 per million tokens. Part of our AI Models & Companies reviews.
Magistral Medium is the model behind the model. When Mistral AI released its first reasoning products on June 10, 2025, the headline for much of the AI community was Magistral Small — 24B parameters, Apache 2.0, runs on a single RTX 4090. The open-weight story travels faster.
But Magistral Medium is how Magistral Small learned to reason. The relationship is structural: Magistral Small was trained through supervised fine-tuning on reasoning traces generated by Magistral Medium, then refined with reinforcement learning from verifiable rewards. Medium is the teacher. Small is the student.
That makes Medium worth understanding on its own terms — not just as the API option for users who do not need local deployment, but as Mistral’s first-ever entry into the reasoning model tier, and the model whose chain-of-thought outputs seeded an open-weight reasoning model that any developer can download.
Background: June 2025 and the Reasoning Model Race
By mid-2025, reasoning models had moved from experimental to expected. OpenAI’s o1 (September 2024) had established that chain-of-thought reasoning at inference time — letting the model “think” before answering — produced meaningfully better results on hard problems than standard instruction-tuned models. DeepSeek R1 (January 2025) showed that reasoning capability could be distilled into open-weight models and released under MIT. QwQ-32B followed. Google’s Gemini 2.5 Pro launched with reasoning built in.
Mistral had not yet released a reasoning model. The company was known for efficient open-weight instruction models — Mistral 7B, Mixtral 8x7B, Mistral Large — but had not entered the chain-of-thought tier. June 10, 2025 ended that absence, with a pair of models announced simultaneously.
Magistral Medium is the larger, proprietary member of that pair. Its role in Mistral’s product architecture is to provide reasoning capability through the API, without requiring local compute or open weights. Magistral Small is the option for developers who need weights they can own and run.
The technical paper covering both models, arXiv:2506.10910, was published June 12, 2025.
What Magistral Medium Is
Magistral Medium is a reasoning model — a model that generates chain-of-thought reasoning before producing its final answer. The model “thinks” through a problem in intermediate tokens that are visible in the output, producing a reasoning trace before the final response.
Mistral has not disclosed the parameter count. The model is available exclusively via API — no downloadable weights, no local deployment. This is a departure from Mistral’s standard practice of releasing open-weight versions alongside proprietary products, and a deliberate business decision: Medium is the commercial API offering, while Small is the open alternative.
Key specifications:
| Property | Value |
|---|---|
| Release date | June 10, 2025 |
| Parameters | Undisclosed |
| Context window | 40,960 tokens |
| Architecture | Reasoning model (chain-of-thought) |
| Weights | Not available |
| API model ID | magistral-medium-2506 |
| Pricing | $2.00/M input, $5.00/M output |
| Platforms | La Plateforme, Amazon SageMaker |
| Languages | English, French, Spanish, German, Italian, Arabic, Russian, Simplified Chinese |
The 40K context window is worth noting. This is shorter than Mistral’s instruction models (Mistral Large 2 offers 128K; Mistral Medium 3.x offers 128K), but it reflects the cost structure of reasoning models. Chain-of-thought generation produces long intermediate outputs — thinking tokens count against the context window. A 40K effective window for a reasoning model means roughly as much usable context as a much larger window on a non-reasoning model, once you account for the space the reasoning trace itself occupies.
The Teacher–Student Relationship with Magistral Small
The most distinctive fact about Magistral Medium is not its benchmark performance — it is what it made possible.
Magistral Small was trained in two stages:
- Supervised fine-tuning (SFT) on reasoning traces generated by Magistral Medium. The Medium model was used to produce high-quality chain-of-thought examples, which were then used to fine-tune Mistral Small 3.1 into a reasoning model.
- Reinforcement Learning with Verifiable Rewards (RLVR) applied on top of SFT. The student model was further refined by rewarding correct final answers on verifiable problems.
This makes Magistral Medium a synthetic data generator as much as a production API model. Its reasoning traces are the raw material from which an open-weight reasoning model was constructed. Users who have deployed Magistral Small are, in a real sense, benefiting from Magistral Medium’s outputs — just compressed and refined into a 24B model they can run locally.
This is the same distillation pattern that produced DeepSeek R1-Distill models from DeepSeek R1, and that underlies much of the reasoning model ecosystem as of 2025. Magistral is Mistral’s implementation of the same knowledge transfer.
Benchmarks
Magistral Medium’s official benchmark results at launch:
| Benchmark | Score | Notes |
|---|---|---|
| AIME 2024 pass@1 | 73.6% | Standard single-sample evaluation |
| AIME 2024 majority@64 | 90.0% | Best-of-64 samples consensus |
| AIME 2025 | 64.9% | Harder problems from the 2025 competition |
| GPQA Diamond | 70.83% | Graduate-level science reasoning |
Benchmark Context
AIME 2024 pass@1 (73.6%): The American Invitational Mathematics Examination pass@1 score represents single-sample performance — what the model gets right on its first try. 73.6% is above Magistral Small’s 70.68% (expected, given the presumably larger parameter count) and above Mistral’s base instruction models. It sits below the reasoning model leaders at launch: DeepSeek R1 was around 79.8%, QwQ-32B around 79.5%, and OpenAI o1 approximately 83.3%.
The Mistral release notes report this represents approximately a 50% relative improvement over a base Mistral Medium model without reasoning training — the chain-of-thought approach meaningfully changes what the underlying architecture can accomplish on math problems.
AIME 2024 majority@64 (90.0%): Best-of-64 sampling — running the model 64 times and taking the consensus answer — reaches 90.0%. This is the metric that shows the model’s theoretical ceiling with compute scaling. Mistral specifically cites this as “on par with DeepSeek-R1-Zero” — DeepSeek’s RL-only reasoning model (before distillation). Whether majority@64 is a practically useful metric depends heavily on the application: few production deployments can afford to run 64 inference passes per query.
AIME 2025 (64.9%): AIME 2025 problems are harder than 2024. The 8.7-point drop from 73.6% to 64.9% between the two years is typical for reasoning models — AIME 2025 was designed to reduce score inflation on models that had over-indexed on 2024 training data.
GPQA Diamond (70.83%): Graduate-level science questions in physics, chemistry, and biology. 70.83% is slightly above Magistral Small’s 68.18%. The gap between the two models is smaller here than on AIME, which is consistent with GPQA requiring less raw mathematical reasoning and more breadth of scientific knowledge.
Pricing and Competitive Position at Launch
At release, Magistral Medium was priced at $2.00/M input, $5.00/M output.
The June 2025 reasoning model pricing landscape for comparison:
| Model | Input | Output | AIME 2024 pass@1 |
|---|---|---|---|
| OpenAI o3-mini | $1.10/M | $4.40/M | ~79.3% |
| DeepSeek R1 API | $0.55/M | $2.19/M | ~79.8% |
| Magistral Medium | $2.00/M | $5.00/M | 73.6% |
| OpenAI o1 | $15.00/M | $60.00/M | ~83.3% |
Magistral Medium costs more per token than both DeepSeek R1 and o3-mini while scoring below both on AIME. That is an honest assessment of its competitive position at launch on the math reasoning benchmark most visible to developers.
The case for Magistral Medium in June 2025 was not benchmark superiority — it was the Mistral ecosystem. Organizations already using Mistral’s API, with data residency requirements or procurement relationships with La Plateforme, gained a reasoning model that integrated directly with their existing infrastructure. The Le Chat consumer interface offered “Flash Answers” mode — Mistral’s claim of up to 10× faster throughput than competitor reasoning models — for users who wanted lower latency at the cost of some reasoning depth.
Flash Answers: Mistral’s Speed Claim
Magistral Medium in Le Chat supports a “Flash Answers” mode alongside standard reasoning. The Flash mode reduces the length of the chain-of-thought trace, trading thoroughness for response speed. Mistral claimed at launch that Flash Answers could deliver responses up to 10× faster than competing reasoning model implementations.
This is a consumer-facing feature, not an API parameter. For API users, the model always runs standard reasoning. Flash Answers is Mistral’s solution to the common complaint about reasoning models: they are slow because the intermediate thinking tokens take time to generate before the final answer appears.
The 10× figure is a marketing claim and reflects throughput on Mistral’s infrastructure compared to competitor hosting — not a fundamental difference in model architecture. Reasoning model speed is heavily infrastructure-dependent.
Multilingual Reasoning
Magistral Medium supports chain-of-thought reasoning in 8 languages: English, French, Spanish, German, Italian, Arabic, Russian, and Simplified Chinese. This is narrower than Magistral Small’s 25+ language support — likely because the teacher model’s multilingual training was more constrained, and the SFT traces that trained Magistral Small included a broader language set.
The eight languages Medium supports cover a substantial portion of the business use cases for API-deployed reasoning: European enterprise use (French, German, Spanish, Italian), Russian-language markets, Arabic-language markets, and Chinese enterprise users. For the majority of the global API market, this is sufficient coverage.
Availability
La Plateforme: The primary access point. Model ID: magistral-medium-2506. Available under Mistral’s standard API terms.
Amazon SageMaker: Available at launch for AWS customers, supporting existing AWS infrastructure integrations.
Le Chat: Preview access for consumer users, with Flash Answers mode.
IBM WatsonX, Azure AI, Google Cloud: Listed as “coming soon” at launch (June 2025).
What Happened After Launch
Magistral Medium’s relevance as a standalone product narrowed significantly over subsequent months.
Mistral Medium 3.5 (released April 29, 2026) effectively superseded it. Mistral Medium 3.5 is a 128B dense open-weight model released under a modified MIT license. It consolidates three previously separate products — Magistral Medium’s reasoning capability, Devstral 2’s coding focus, and Mistral Medium 3.1’s instruction following — into a single model with a reasoning toggle. Pricing: $1.50/M input tokens flat. SWE-bench Verified: 77.6%.
At $1.50/M (versus Magistral Medium’s $2.00/M input, $5.00/M output), with open weights and a broader capability set, Mistral Medium 3.5 is the current recommended option for new API deployments requiring reasoning from Mistral’s lineup. Magistral Medium’s model ID (magistral-medium-2506) remains active on La Plateforme as of May 2026, but Mistral’s documentation routes new users toward Medium 3.5.
The evolution is consistent with how reasoning capability tends to commoditize: specialized reasoning models get rolled into general-purpose models as the capability becomes standard rather than differentiated.
Comparison to Contemporaries
vs. Magistral Small (released same day)
The 3-point AIME gap (73.6% vs 70.68%) and 2.65-point GPQA gap (70.83% vs 68.18%) between Medium and Small are real but modest. For most applications, the difference between the two models is not the benchmarks — it is the deployment model. Small runs locally under Apache 2.0; Medium requires the Mistral API. Developers who can tolerate API dependency and need slightly better math performance have Medium available. Everyone else has Small.
vs. OpenAI o3-mini (January 2025)
o3-mini at $1.10/M input, $4.40/M output scores approximately 79.3% on AIME 2024 — 5.7 points above Magistral Medium at lower cost. o3-mini was already available five months before Magistral launched. From a pure benchmark-per-dollar standpoint, o3-mini was a stronger choice at the time.
vs. DeepSeek R1 API
DeepSeek R1 at $0.55/M input, $2.19/M output scores approximately 79.8% on AIME — 6.2 points above Magistral Medium at significantly lower cost. The case for Magistral Medium over R1 is ecosystem, not economics.
vs. Mistral Medium 3.5 (April 2026)
Mistral Medium 3.5 is cheaper ($1.50/M input flat), open-weight, broader (reasoning + coding + instruction), and scores 77.6% on SWE-bench. For any new deployment, Medium 3.5 is the current choice within Mistral’s lineup.
Strengths and Weaknesses
Strengths:
- Historical significance: Magistral Medium was Mistral AI’s first reasoning model — the entry point into a new capability tier for the company
- Teacher model: Generated the reasoning traces that trained Magistral Small; understanding Medium is understanding how Small’s capability was built
- Multilingual reasoning: 8-language chain-of-thought support including Arabic and Russian — broader than most enterprise-grade API reasoning models at launch
- AIME majority@64 (90.0%): With sufficient compute, approaches DeepSeek-R1-Zero level math performance
- API integration: Slots directly into existing Mistral API deployments without infrastructure changes
Weaknesses:
- Benchmark underperformance vs. price: o3-mini and DeepSeek R1 both outscored Magistral Medium on AIME at lower cost in June 2025
- No open weights: Unlike Magistral Small, there is no local deployment path — full API dependency
- Parameter count undisclosed: Unusual for a model this visible; makes architectural comparison difficult
- 40K context ceiling: Shorter than Mistral’s instruction models; constrains long-document and multi-turn reasoning use cases
- Superseded: Mistral Medium 3.5 (April 2026) is the current recommended option at lower cost and with open weights
Rating: 3.5/5
Magistral Medium earns its place in Mistral’s product history as the model that proved Mistral could build in the reasoning tier — and, more concretely, as the teacher model whose outputs enabled Magistral Small.
The half-point deduction from Magistral Small’s rating reflects the pricing reality at launch: $2.00/M input for 73.6% AIME performance was not the best ratio available when o3-mini and DeepSeek R1 were already on the market. Magistral Medium’s value was always primarily for Mistral’s existing API ecosystem, not for developers optimizing benchmark-per-dollar.
By April 2026, Mistral Medium 3.5 made the pricing calculus straightforward — lower cost, open weights, broader capability. Magistral Medium’s active story lasted roughly ten months as a distinct product.
For the record: on June 10, 2025, Magistral Medium was Mistral AI’s best reasoning model. It taught its smaller counterpart how to think. That contribution persists in every Magistral Small deployment running today.
This review is part of ChatForest’s ongoing coverage of the AI model landscape. See our Magistral Small review for the open-weight companion model, our Mistral AI company overview for context on Mistral’s full lineup, and our AI model reviews index for comparisons across other providers.