Mistral AI
The best prompting documentation of any provider here, an EU inference region, and almost the whole line-up available as open weights.
How to prompt it
Open with a role and task in one line: you are a given role, and your task is a given task. Mistral documents this exact shape.
Use Markdown or XML-style tags — Mistral explicitly blesses both, on the grounds that they are readable, parsable and familiar. It is the only provider here that endorses both equally.
Replace vague quantities with measurements. Mistral names “too long”, “too short”, “many” and “few” as words to avoid, and asks for objective measures instead.
Never ask it to count words. Mistral says this outright; give it character counts as input instead.
Use a worded scale rather than a numeric one when you want a rating — very low through very good beats one to five.
Decide where durable instructions go by testing. Mistral is unusual in documenting that a system message and a user message are interchangeable, and telling you to try both.
Strengths
- It documents anti-patterns, not just good practice — what not to ask for, and why. Almost no other vendor does this.
- Nearly the entire line-up ships as open weights under Apache 2.0, including the largest model. You can self-host the same model you prompt through the API.
- A genuine EU inference endpoint, which matters for anyone weighing data residency.
- Assistant prefill is supported, so you can force a reply to start in a specific format — one of the strongest levers for controlling output shape.
Weaknesses
- Reasoning control is binary: on or off, with no middle setting. You cannot buy a little more thinking the way you can elsewhere.
- Specifications are published inconsistently. Maximum output is not stated anywhere, and context windows appear only on individual model cards rather than in any comparison.
- The EU residency story has caveats Mistral publishes itself: a price premium, several features unavailable on regional endpoints, and account and billing metadata that may still be handled outside the chosen region.
- One mid-tier model carries a modified licence with a revenue-based carve-out rather than a clean open one, and the threshold is not stated on the model card.
Models
Mistral Large 3
State-of-the-art, open-weight, general-purpose multimodal model.
- Context
- 256K
- Max output
- —
Good for: General work where you want frontier-class output at a low price, or the option to self-host the same model.
Watch out: Maximum output is not published, so size long generations by testing.
Mistral Medium 3.5
Frontier-class multimodal model for agentic and coding use.
- Context
- 256K
- Max output
- —
Good for: Agentic and coding workloads, with documented strong adherence to system prompts.
Watch out: Much more expensive per output token than the larger model, and its licence is modified rather than fully open.
Mistral Small 4
Hybrid model unifying instruct, reasoning and coding in one.
- Context
- 256K
- Max output
- —
Good for: The value pick, and the one most people should start with.
Watch out: Reasoning is either off or high; there is no middle setting.
Ministral 3 8B
Efficient small model with text and vision.
- Context
- 256K
- Max output
- —
Good for: High volume, low cost, and running locally — it is about 5 GB quantised.
Watch out: Give it more explicit instruction than you would the larger models.
Sources and grounding
- 1Prompt engineering best practicesMistral AISource for every prompting rule on this page, including the Markdown-or-XML statement and the anti-patterns.
- 2ModelsMistral AILineup, positioning and licence per model.
- 3Regional inferenceMistral AIThe EU endpoint, its price premium and the metadata caveat.