Max-output mapping
How the provider derives a model’s maximum output tokens. Mistral’s API returns no per-model max-output field, but it enforces a per-model completion-token ceiling that the model pages do not document: a request with a larger max_tokens fails with 400 invalid_request_max_tokens (code 3058).
The mapping, checked most-specific-first:
- Documented output limit — for the models where Mistral documents an explicit output token limit
- Wire-enforced completion ceiling — for models whose ceiling was probed on the live API
- Conservative fallback (262144, capped at the context window) — everything else
A value above the true ceiling hard-fails requests with 400 code 3058; a conservative one merely caps output.
| Model family | Documented output limit | Context |
|---|---|---|
| zai-glm | 131072 | 1M |
| deepseek | 131072 | 1M |
| Model family | Wire-enforced completion ceiling | Context |
|---|---|---|
| mistral-large-4 | 262144 | 1M |