Overview
Mistral Medium (2023) was the top tier of Mistral AI's first commercial API, La Plateforme, announced on December 11, 2023. Where the launch also exposed mistral-tiny (Mistral 7B) and mistral-small (Mixtral 8x7B) as endpoints over open-weight models, the mistral-medium endpoint served an undisclosed internal "prototype model" that Mistral was testing in a deployed setting. It was the company's first genuinely closed, pay-only flagship.
The model handles text only, across English, French, Italian, German and Spanish plus code, with a 32K-token context window. Mistral never published its parameter count or architecture, and the only first-party benchmark figure released at launch was an MT-Bench score of 8.6, ahead of the mistral-small (Mixtral 8x7B) endpoint's 8.3. It was positioned as Mistral's most capable hosted model until Mistral Large arrived in February 2024.
Mistral Medium (2023) was later dated and tagged as mistral-medium-2312 to distinguish it from successors. Mistral deprecated the model on November 30, 2024 and retired it from the API on June 16, 2025; the modern replacement is the unrelated Mistral Medium 3 (2025). This page documents the original 2023 line for historical reference - it is no longer callable.
| Released | 2023-12-11 |
|---|---|
| License | Proprietary (commercial API only) |
| Weights | API only |
| Parameters | Undisclosed (Mistral never published the model size) |
| Context | 32K tokens |
| Architecture | Undisclosed proprietary transformer. Mistral described the endpoint as serving an internal "prototype model" it was testing in a deployed setting, distinct from its open-weight Mistral 7B and Mixtral 8x7B releases. |
| Knowledge cutoff | Not disclosed by Mistral AI |
| Modalities | text-in, text-out |
| Status | Retired |
Benchmarks
- MT-Bench8.6%
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Pricing
| Input | $2.75 per 1M tokens |
|---|---|
| Output | $8.10 per 1M tokens |
Mistral did not publish per-token pricing in the original launch post; these are the historical mistral-medium-2312 API rates as recorded by Artificial Analysis. The endpoint is now retired and no longer billable.
Strengths
- First closed Mistral flagship: stronger than the open-weight mistral-small (Mixtral 8x7B) endpoint at launch (MT-Bench 8.6 vs 8.3)
- Native multilingual coverage of English, French, Italian, German and Spanish, plus code
- 32K-token context window, generous for a late-2023 hosted model
- Available through a simple OpenAI-compatible REST API on Mistral's La Plateforme from day one
Best for
- General-purpose chat and instruction following in English and major European languages (historical)
- Multilingual drafting, summarization and translation across French, Italian, German and Spanish
- Code generation and explanation tasks where an open-weight model was not preferred
- A 2023-era cloud alternative to other proprietary chat APIs for European teams
How to access
| Provider | Model ID |
|---|---|
| Mistral AI (La Plateforme) ↗ | mistral-medium-2312 |
Mistral Medium — every version
The full lineage of the Mistral Medium line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.
| Version | Released | Context | License |
|---|---|---|---|
| Mistral Medium 3.5current | 2026-04-28 | — | MIT |
| Mistral Medium 3.1 | 2025-08-12 | — | Open weights |
| Mistral Medium 3 | 2025-05-07 | — | Open weights |
| Mistral Medium (2023) | 2023-12 | — | Proprietary |
FAQ
Is Mistral Medium (2023) still available?
No. The original mistral-medium endpoint, later tagged mistral-medium-2312, was deprecated on November 30, 2024 and fully retired from Mistral's API on June 16, 2025. It can no longer be called.
Is Mistral Medium (2023) the same as Mistral Medium 3?
No. Mistral Medium 3 is a separate, much newer model released in 2025. The 2023 line was an early prototype-backed endpoint and shares only the name with the current Medium series.
What was special about the mistral-medium endpoint at launch?
It was Mistral AI's first closed, pay-only flagship. Alongside it, mistral-tiny and mistral-small simply wrapped the open-weight Mistral 7B and Mixtral 8x7B models, but mistral-medium served an undisclosed internal prototype with higher performance (MT-Bench 8.6).
Were the weights for Mistral Medium (2023) ever released?
No. Unlike Mistral 7B and Mixtral 8x7B, the Mistral Medium prototype was never open-sourced. Its parameter count, architecture and training data were never disclosed.