Feature request: `response_format` support for MiniMax M2.5 via OpenAI-compatible API

[joneslloyd]

Summary

MiniMax M2.5 does not support response_format (neither json_object nor json_schema) via the OpenAI-compatible chat completions endpoint. The native API documentation explicitly states this feature is "only supported by MiniMax-Text-01". This makes M2.5 unsuitable for production workloads that require structured JSON output.

Current behaviour

When sending response_format: { "type": "json_object" } to POST /v1/chat/completions with M2.5:

• The parameter is silently ignored. No error or warning is returned.
• The model generates free-form text that may contain JSON, but is not constrained to valid JSON.
  Additionally, when relying on prompt-based JSON instructions (the only available workaround), M2.5 frequently produces structurally malformed JSON. We observed a recurring defect across multiple production runs: the model drops the opening { brace for object elements inside arrays:

// Expected:
[{title: SEO Improvement, score: 8}, {title: Content Gap, score: 6}]

// Actual M2.5 output:
[{title: SEO Improvement, score: 8}, title: Content Gap, score: 6}]

This occurred consistently with schemas containing 10+ fields with nested objects and arrays. We had to implement a context-aware JSON repair function to work around this, which adds complexity and is inherently fragile.

Expected behaviour

response_format should be supported for M2.5 via the OpenAI-compatible API:

1. { "type": "json_object" } should constrain output to syntactically valid JSON
2. { "type": "json_schema", "json_schema": { "schema": ... } } should constrain output to JSON matching the provided schema
   If full json_schema mode is not feasible for M2.5, json_object mode alone would be a significant improvement; it would guarantee valid JSON syntax even without schema conformance.

Documentation gap

The OpenAI-compatible API documentation page does not mention response_format at all; it is absent from the supported parameters table and not listed among the "ignored" parameters either. The only documentation of response_format is in the native API spec, which states it is MiniMax-Text-01 only.
A clear compatibility matrix (per model, per API flavour) would help developers evaluate MiniMax models before committing to integration.

Impact

We evaluated M2.5 for a production AI pipeline processing 12 agents with structured output requirements. Despite M2.5's strong raw capabilities (204K context, competitive pricing), we abandoned it after 2 production runs due to the combination of:

1. No response_format support (silent ignore)
2. Consistent malformed JSON output from prompt-based workarounds
3. Error 2013 ("invalid chat setting") when using multiple system messages (which blocked one workaround path for schema injection)
   We switched to OpenAI GPT-5-nano with native response_format support (json_schema mode with strict: true), which works reliably.

Questions

1. Is there a roadmap for adding response_format to M2.5 via the OpenAI-compatible API?
2. Is the malformed JSON issue (dropped opening braces) a known model-level problem?
3. Would the MiniMax team consider adding response_format to the OpenAI-compatible endpoint for M2.5, even if only json_object mode?