Mistral
France's frontier lab, with the small models everyone self-hosts.
Mistral is the main model family from the French lab of the same name, and it splits in a way worth knowing about: the Small models publish their weights so you can download and run them, while Medium and Large are sold through the API only. Small 4 folds several older models into one and is cheap enough to use by default; Medium 3.5 is the one to reach for on agentic and coding work. Mistral's other lines are separate products with their own names — Devstral and Codestral for code, Ministral for tiny on-device models, and Voxtral for speech, which has its own entry here.
At a glance — Mistral Small 3.2 24B
- Cost
- $0.094 in · $0.25 out per 1M tokensBudget
- Memory
- 256K tokens — about 400 pages
- Reads
- Text · Images
- Can do
- Use tools (agents)
- Structured output
- Thinking mode (not supported)
- Reliability
- 100.00% uptime, last 24h
- Released
- Jun 20, 2025
Advanced details
- API model ID
mistralai/mistral-small-3.2-24b-instruct- Max output
- 16K tokens
- Knowledge cutoff
- 2023-10-31
- Tokenizer
- Mistral
Providers
The same model, hosted by different companies. OpenRouter picks one per request and falls back to the next if it fails.
| Provider | Input /1M | Output /1M | Cached /1M | Max output | Uptime 24h |
|---|---|---|---|---|---|
| DeepInfra · fp8 | $0.075 | $0.20 | — | 16K | 99.87% |
| Parasail · bf16 | $0.09 | $0.30 | $0.05 | 33K | 99.68% |
| Venice · fp8 | $0.094 | $0.25 | — | 16K | 99.67% |
| Mistral · eu | $0.10 | $0.30 | $0.01 | 26K | 100.00% |
Supported parameters
- frequency_penalty
- logit_bias
- logprobs
- max_tokens
- min_p
- presence_penalty
- repetition_penalty
- response_format
- seed
- stop
- structured_outputs
- temperature
- tool_choice
- tools
- top_k
- top_logprobs
- top_p
Live from OpenRouter, refreshed hourly.
Key info
- Pricing
- Freemium
- Category
- Models
- Best for
- Intermediate
- Made by
- Mistral AI, in France
- Weights
- Open for Small and Nemo, closed for Medium and Large
- Good for
- A cheap default model, and EU-hosted inference