Devstral Medium
devstral-medium-2507
Devstral Medium is a text model from 米斯特拉尔AI. The snapshot records a release date of 2025-07-10. Context window: 128K tokens. Maximum output: 128K tokens. Input: text. Output: text. Capabilities recorded: tool calls.
Directory snapshot 2026-10-26 · Chinese dossier: 完整中文页
Official list price
CNY / 1M TOKENSOfficial list price in the directory snapshot: ¥2.67 input and ¥13.37 output per 1M tokens, in CNY. Relay prices below are what individual sites listed; they can be lower or higher, and they are not official prices.
Relay coverage
NONE IN SNAPSHOTNone of the sites in our current public snapshot listed this model ID. That does not mean no relay offers it; it means we have no listing to show. If you find one, the detection page can test it.
How to verify this model on a relay
THREE STEPS- Send a short request to the relay and read the
modelfield in the response. It must match the ID you sent; a different ID is a mapping, not necessarily a substitution. - Run the same prompt against a known-good endpoint and compare prompt-token counts. A large, repeatable gap points at a different tokenizer or a different backend.
- Run the seven checks. Billing, context retention, error shape, parameter compliance, tools, and streaming each produce evidence you can quote.
Same publisher
米斯特拉尔AIQuestions
FAQWhat is the context window of Devstral Medium?
The snapshot records 128,000 tokens.
How much does Devstral Medium cost?
Official list price: ¥2.67 input / ¥13.37 output per 1M tokens in CNY, per the directory snapshot. No relay price is recorded.
Is a relay really serving Devstral Medium?
A model ID in a response is not proof. The seven-check detection compares tokenizer ratios, billing, error shape, parameter compliance, tools, and streaming against the advertised family. Run it on a key you can afford to spend, and treat one run as a spot check, not 24/7 monitoring.
Specifications and prices come from the 2026-10-26 public directory snapshot. We do not publish IQ scores or structured-output claims for models where we have no reliable data.