Llama 4 JSON Tool Calling: Strict Array Format

Llama 4 also supports a JSON zero-shot function calling format where the assistant outputs a JSON array of objects specifying name and parameters, followed by the same <|header_start|>ipython<|header_end|> response turn.

<|header_start|>assistant<|header_end|> [ { "name": "get_weather", "parameters": { "city": "San Francisco" } }, { "name": "get_weather", "parameters": { "city": "Seattle" } } ]<|eot|> <|header_start|>ipython<|header_end|> [ { "response": "Sunny 75" }, { "response": "Rainy 65" } ]<|eot|> <|header_start|>assistant<|header_end|>