Provider notes
Per-provider and per-model quirks to be aware of.
These are documented behaviors that differ from the generic OpenAI request shape.
Kimi (kimi/kimi-k2.6)
kimi-k2.6 ignores temperature, top_p, and penalty parameters.
MiniMax
MiniMax models ignore presence_penalty and frequency_penalty parameters.
Bedrock — kimi-k2-thinking
bedrock/kimi-k2-thinking is a reasoning model that uses an internal thinking budget. Use max_tokens ≥ 1000 to ensure output is produced.
Bedrock — gpt-oss-20b
bedrock/gpt-oss-20b requires max_tokens ≥ 500 to produce output. Accepts reasoning_effort.
Azure — gpt-5.x
azure/gpt-5.5 and azure/gpt-5.6-sol are reasoning models. The gateway translates max_tokens to Azure's max_completion_tokens for you, but temperature, top_p, penalty params, and logprobs are silently dropped — only default sampling is supported.
thinking and reasoning_effort are model-specific
These are not gateway-wide flags, and the gateway does not reject them with an error. Support is per-model:
bedrock/claude-*acceptsthinking; reasoning content comes back as areasoningfield (see Anthropic (Claude) models).bedrock/gpt-oss-*andazure/gpt-5.*acceptreasoning_effort.azure/gpt-5.*silently dropsthinkingif sent.- Every other provider forwards these fields as-is to its upstream API, which may ignore or error on them depending on that provider's own request validation.
Omit both fields unless you're targeting one of the reasoning models above.