Tencent: Hy4 preview

Supported parameters

• frequency_penalty

  Frequency penalty

  Lowers the chance of a token in proportion to how often it already appeared.

• include_reasoning

  Include reasoning

  Include a reasoning summary in the response.

• logit_bias

  Logit bias

  Add or subtract from token ids to encourage or ban them.

• logprobs

  Logprobs

  Return the log probability of each output token.

• max_completion_tokens

  Max completion tokens

  Same idea as max_tokens. Some newer models use this field instead.

• max_tokens

  Max tokens

  Upper bound on generated tokens. Output stops at the limit.

• min_p

  Min-p

  Drops tokens below this probability threshold.

• presence_penalty

  Presence penalty

  Lowers the chance of any token that has already appeared.

• reasoning

  Reasoning

  Enable or configure reasoning. The fields differ by model.

• reasoning_effort

  Reasoning effort

  How much compute to spend reasoning. Often low, medium, or high.

• repetition_penalty

  Repetition penalty

  Penalizes text that already appeared. Common on open models.

• response_format

  Response format

  For example json_object or json_schema, to constrain the shape.

• seed

  Seed

  A fixed seed helps reproduce sampling. It does not guarantee byte-identical output.

• stop

  Stop

  Stop when these strings are produced. A string or an array of strings.

• structured_outputs

  Structured outputs

  Generate a parseable structure that follows a JSON Schema.

• temperature

  Temperature

  Sampling temperature. Lower is steadier, higher is more varied. Often 0 to 2.

• tool_choice

  Tool choice

  Whether to call a tool: auto, none, or a named function.

• tools

  Tools

  Function definitions the model may call via tool_calls.

• top_k

  Top-K

  Sample only from the K most likely tokens.

• top_logprobs

  Top logprobs

  Also return the most likely alternatives at each position.

• top_p

  Top-p

  Sample only from tokens whose cumulative probability reaches top_p.