Mistral Large 4 function calling: validate arguments against Sume

You can use Mistral Large 4 to draft arguments for a Sume call, but check them before they spend anything. Mistral's model page marks it Public Preview as of October 6, 2026, open-weight, with a 1M-token context and function calling and structured outputs among its capabilities. A preview model can produce a well-formed but wrong argument, so the contract check belongs in your code.

What the page lists

Only the facts relevant to a tool loop are shown.

The Sume side of the check

Sume's hosted MCP has tools_schema, which returns one tool contract by name. Fetch the contract for the tool the model chose, validate the model's arguments with a JSON Schema library, and reject on mismatch. Then call with a fresh idempotency_key and dry_run=true first, as the tools and gates page recommends before paid generation.

This post covers calling Mistral from your own code. It does not say which model Sume uses in its own agent features; see Sume's docs for that.

The validation loop

The order is: model proposes a tool name and arguments; your code calls tools_schema for that name; you validate; on failure you send the validation error back once and let the model retry; on a second failure you stop. Only a valid call goes on to dry_run=true, then to the real call.

Preview models change. Pin the id mistral-large-4 in config, log it with each call, and re-run your examples when Mistral updates the page.

• Validate before any paid call.
• One retry on a schema error.
• Log the model id.