Mistral
Prompting Mistral
How to prompt Mistral models across Vibe, Studio/API, open-weight variants, coding, OCR, and agentic workflows.
What this model is particularly good at:
- Broad model lineup spanning frontier multimodal, efficient hybrid, OCR, audio, moderation, code, and open-weight models
- Useful European/commercial alternative with strong API docs and a first-party prompting guide
- Clear distinction between Vibe product behavior and Studio/API model selection
Current model routing
| Model | Status | Best fit | API ID |
|---|---|---|---|
| Mistral Medium 3.5 | default | Frontier-class multimodal work optimized for agentic and coding use cases | Varies by surface |
| Mistral Small 4 | current | Efficient hybrid instruct, reasoning, and coding in one model | Varies by surface |
| Mistral Large 3 | current | Open-weight general-purpose multimodal deployment | Varies by surface |
| Ministral 3 (14B / 8B / 3B) | open-weight | Small open-weight text and vision models (Apache 2.0) for local and edge deployment | Varies by surface |
| Mistral OCR 4.1 | specialized | Document intelligence and OCR workflows | Varies by surface |
| Codestral / Leanstral / Voxtral | specialized | Code completion, formal proof work, transcription, and voice tasks | Varies by surface |
Which model to choose
Use Mistral Medium 3.5 for the strongest general Mistral work, especially agentic and coding tasks. Use Mistral Small 4 when you want an efficient hybrid model that unifies instruct, reasoning, and coding. Use Mistral Large 3 when open-weight deployment matters. Route OCR, audio, code completion, and formal proof work to the specialized model families instead of trying to make one chat prompt do every job.
Prompting in Vibe
In Mistral’s GUI, prompt the workflow, not just the model. Le Chat is now Vibe, a single product with three modes: Vibe Work for multi-step productivity and agentic work across apps and tools, Vibe Code for developer coding (CLI, VS Code extension, and web), and Vibe Chat for turn-based conversation. Pick the mode that matches the job.
Use concise instructions, define the output format, and give examples only where they remove ambiguity. If the product exposes work/code modes or connectors, use those controls rather than describing them indirectly in prose.
Prompting through the API
Mistral’s prompting docs preserve the classic separation between system and user prompts: use the system prompt for durable behavior and the user prompt for the actual request. The model overview should drive model routing before prompt tuning starts.
For document/OCR workflows, send the document through the document/OCR product surface when available rather than pasting lossy OCR text into a general chat prompt.
Current prompting guidance
- Set system-level behavior separately from the user task.
- Choose the model family by task: general, code, OCR, audio, moderation, or formal proof.
- Keep prompts clear and direct.
- Use examples for classification or formatting boundaries.
- For open-weight Mistral deployments, verify the exact model card and chat template.
Examples
GUI prompt:
Review this product spec and identify which requirements should be handled by OCR,
which by a general reasoning model, and which need human review. Keep the result as
a three-column table.API request skeleton:
{
"model": "mistral-medium-3.5",
"messages": [
{
"role": "system",
"content": "You are a concise contract-review assistant. Flag only concrete obligations, deadlines, and risks."
},
{
"role": "user",
"content": "Extract the obligations from this contract excerpt and return a table with owner, obligation, deadline, and source sentence.\n\n{{contract_excerpt}}"
}
]
}Prompt upgrade example:
Stale pattern: “Use the biggest Mistral model and solve everything in one response.”
Better pattern: “Route OCR, coding, and audio to the matching Mistral model or product surface before tuning the prompt.”