Explore DeepSeek-V3

The conversation sent to the chat model in chronological order. Each JSON item contains a role and its text content.

Defines functions the model is allowed to request, including each function’s name, purpose, and JSON parameter schema. Supplying a tool does not execute it—the application must run the requested function and return its result to the model.

Sends small chunks as the model generates them, allowing the interface to display the beginning of a response sooner. It changes delivery timing rather than model quality; disable it when the client needs one complete response object.

Adjusts how strongly the model favors its most likely next token. Lower values produce more focused and repeatable output, while higher values allow less-likely choices and increase variety; it only has an effect when sampling is enabled.

Limits sampling to the smallest set of likely tokens whose probabilities add up to this value. Lower settings make output more focused, while a value near 1 keeps more alternatives available; it only affects sampled generation.

Caps the length of the generated response in tokens, which are pieces of words rather than whole words. The model can stop naturally before the limit, but a value that is too low may cut off the answer and a high value can increase latency and cost.

Reduces the score of a token more each time it has already appeared in the response. Positive values discourage repeated words and phrases, zero makes no adjustment, and negative values encourage repetition.

Reduces the score of any token once it has appeared, regardless of how many times it was used. Positive values encourage new vocabulary or topics, zero makes no adjustment, and negative values favor staying with existing wording.

Chooses between an unrestricted natural-language response and a syntactically valid JSON object. When selecting JSON, explicitly describe the desired keys and mention JSON in the prompt so the model knows what structure to produce.

Supplies one or more literal delimiters that stop generation when matched. The match is sensitive to exact spelling, spacing, and capitalization, so use distinctive sequences that are unlikely to appear naturally in the response.

Determines how the model uses the supplied tool definitions. Auto lets it choose between answering and calling a tool, None forbids tool calls, and Required forces a tool call even if the model could answer directly.

Returns the logarithm of the probability assigned to each generated token. Values closer to zero indicate more likely choices; enable this for scoring or debugging, knowing it increases response size and is not a calibrated measure of factual confidence.

Returns this many alternative token candidates and their log probabilities for every output position. Use it to inspect competing word choices or build custom scoring; it normally requires token log probabilities to be enabled.