Mistral names its models by size band, not by generation. Large is the flagship, Medium the workhorse, Small the value tier, and Ministral the edge-sized models built to run on laptops and phones. Around those sit specialists: Codestral for code completion, Devstral for agentic coding, Voxtral for speech, and Nemo, an older small model that is still one of the cheapest things you can rent.
Reading the model id
An id like mistral-small-2603 means the Small family, snapshot from March 2026 (YYMM). ministral-8b-2512 is the 8-billion-parameter Ministral released December 2025. The -latest aliases (for example mistral-small-latest) always point at the newest snapshot, which is convenient for prototypes and dangerous for production, because the price and behaviour can change under you. Pin a dated id once something ships.
| Family | Current model | Input | Output | Context | Pick it for |
|---|---|---|---|---|---|
| Large | Mistral Large 3 2512 | $0.50 | $1.50 | 262K | Best general quality Mistral sells; long documents, hard questions |
| Medium | Mistral Medium 3.5 | $1.50 | $7.50 | 262K | Newest mid-tier; check it earns its premium over Large 3 on your tasks |
| Medium | Mistral Medium 3.1 | $0.40 | $2.00 | 131K | Previous mid-tier, still a strong default for chat and drafting |
| Small | Mistral Small 4 | $0.15 | $0.60 | 262K | Everyday summarising, extraction, classification, open-weight |
| Small | Mistral Small 3.2 24B | $0.075 | $0.20 | 131K | Same job, older snapshot, even cheaper |
| Ministral | Ministral 3 8B 2512 | $0.15 | $0.15 | 262K | Bulk jobs where output volume is high; runs on a laptop |
| Ministral | Ministral 3 3B 2512 | $0.10 | $0.10 | 131K | Tagging, routing, on-device; cheapest new Ministral |
| Codestral | Codestral 2508 | $0.30 | $0.90 | 256K | Autocomplete and fill-in-the-middle code generation |
| Devstral | Devstral 2 2512 | $0.40 | $2.00 | 262K | Agentic coding: multi-file edits, tool use |
| Voxtral | Voxtral Small 24B 2507 | $0.10 | $0.30 | 32K | Speech understanding and transcription-style tasks |
| Nemo | Mistral Nemo | $0.019 | $0.03 | 131K | The cheapest row in the catalogue; simple high-volume text |
Which one should you actually pick?
- Default for most work: Mistral Small 4. It has the full 262K context, it is open-weight, and it is a fraction of the Large price.
- When Small gets it wrong: step up to Mistral Large 3 or Medium 3.1, then test Medium 3.5 only if the others fail and the task is worth five times the output price.
- High-volume, short outputs: Ministral 3 8B or 3B. Symmetric pricing (input and output cost the same) makes them predictable for classification and routing.
- Code in an editor: Codestral. Code agents that edit whole repos: Devstral 2.
- Audio in, text out: Voxtral Small. Watch the 32K context.
- If you are truly cost-bound: Mistral Nemo is one of the cheapest hosted models anywhere. Quality is dated; try it before you scoff.
Context windows
Most current Mistral models sit at 131K or 262K tokens of context, which is roughly a 200- to 400-page book. That is plenty for almost any single document. The catch is that context is not free: every token you put in the window is billed on every request. A 200K-token prompt on Large 3 costs about $0.10 per call before the answer. Pass what the model needs, not the whole archive.
Knowledge check
You need to classify 2 million short support tickets into 12 categories. Which Mistral model is the sensible first choice?