Schemas

Who authored this message. system sets behavior, user asks, assistant answers (or requests tool calls), tool returns a tool result.

The message text. May be null on an assistant message that only carries tool_calls.

Optional participant name to disambiguate same-role messages.

Tool calls requested by the assistant (assistant messages only).

Required when role is tool — the id of the tool call this message responds to.

Only function tools are supported.

Identifier to echo back in the tool role reply as tool_call_id.

A model ID from the current catalog — real provider model names, all MiniMax. The core tiers are minimax-m3 (flagship), minimax-m2.7 (balanced daily driver) and minimax-m2.5 (fast); the -highspeed variants are latency-tuned versions of their standard siblings at roughly 2x the price; minimax-m2.1 and minimax-m2 are older M-series generations; minimax-m2-her is role-play/dialogue chat tuning; minimax-text-01 is the long-context text model. There are no invented aliases; any other ID returns an error. The catalog can change with notice.

The conversation so far — an array of role/content messages, oldest first.

Maximum number of tokens to generate in the completion.

Sampling temperature. Lower is more deterministic, higher is more creative. Alter this or top_p, not both.

Nucleus sampling — only tokens in the top top_p probability mass are considered. Alter this or temperature, not both.

Number of completion choices to generate per request.

Set true for a Server-Sent Events stream of deltas.

Up to 4 sequences where generation stops (string or array).

Penalize tokens by how often they already appeared — positive values reduce verbatim repetition.

Penalize tokens that appeared at all — positive values push toward new topics.

Function tools the model may call. Supported — the gateway passes tool definitions and tool-call turns through to the upstream model.

Controls tool use. auto (the default) and none are honoured by the M-series. required and the named-function object are accepted by the API but ignored by the model, which decides for itself. Tested 2026-07-26. Design for a model that chooses, and always handle a turn with no tool call.

Unix timestamp (seconds).

The model that actually served the request. During a provider incident this may be a smaller model of the same family as the one you requested.

One entry per requested choice (n).

Token usage in Chat Completions shape.

A model ID from the current catalog — real provider model names, all MiniMax. The core tiers are minimax-m3 (flagship), minimax-m2.7 (balanced daily driver) and minimax-m2.5 (fast); the -highspeed variants are latency-tuned versions of their standard siblings at roughly 2x the price; minimax-m2.1 and minimax-m2 are older M-series generations; minimax-m2-her is role-play/dialogue chat tuning; minimax-text-01 is the long-context text model. There are no invented aliases; any other ID returns an error. The catalog can change with notice.

The input — a plain string, or a list of typed input items (e.g. {"role": "user", "content": "..."}).

System-style instructions for the model.

Maximum number of tokens to generate in the response.

Set true for a Server-Sent Events stream of response events.

The model that actually served the request. During a provider incident this may be a smaller model of the same family as the one you requested.

completed, incomplete, or failed.

Output items — for text generation, one assistant message.

Token usage in Responses-API shape.

A model ID from the current catalog — real provider model names, all MiniMax. The core tiers are minimax-m3 (flagship), minimax-m2.7 (balanced daily driver) and minimax-m2.5 (fast); the -highspeed variants are latency-tuned versions of their standard siblings at roughly 2x the price; minimax-m2.1 and minimax-m2 are older M-series generations; minimax-m2-her is role-play/dialogue chat tuning; minimax-text-01 is the long-context text model. There are no invented aliases; any other ID returns an error. The catalog can change with notice.

Maximum number of tokens to generate. Required by the Anthropic Messages format.

Alternating user / assistant turns, oldest first.

Optional system prompt — a top-level field here, not a message role.

Sampling temperature, 0–1 in the Anthropic format. Lower is more deterministic.

Set true for a Server-Sent Events stream of message events.

Custom sequences that stop generation.

The model that actually served the request. During a provider incident this may be a smaller model of the same family as the one you requested.

Content blocks — for text generation, one text block.

end_turn, max_tokens, or stop_sequence.

Which stop_sequences entry matched, if any.