api-inference documentation

Chat Completion

Hugging Face's logo
Join the Hugging Face community

and get access to the augmented documentation experience

to get started

Chat Completion

Generate a response given a list of messages in a conversational context, supporting both conversational Language Models (LLMs) and conversational Vision-Language Models (VLMs). This is a subtask of text-generation and image-text-to-text.

Recommended models

Conversational Large Language Models (LLMs)

Conversational Vision-Language Models (VLMs)

API Playground

For Chat Completion models, we provide an interactive UI Playground for easier testing:

  • Quickly iterate on your prompts from the UI.
  • Set and override system, assistant and user messages.
  • Browse and select models currently available on the Inference API.
  • Compare the output of two models side-by-side.
  • Adjust requests parameters from the UI.
  • Easily switch between UI view and code snippets.

Access the Inference UI Playground and start exploring: https://huggingface.co./playground

Using the API

The API supports:

  • Using the chat completion API compatible with the OpenAI SDK.
  • Using grammars, constraints, and tools.
  • Streaming the output

Code snippet example for conversational LLMs

Python
JavaScript
cURL

Using huggingface_hub:

from huggingface_hub import InferenceClient

client = InferenceClient(api_key="hf_***")

messages = [
	{
		"role": "user",
		"content": "What is the capital of France?"
	}
]

stream = client.chat.completions.create(
    model="google/gemma-2-2b-it", 
	messages=messages, 
	max_tokens=500,
	stream=True
)

for chunk in stream:
    print(chunk.choices[0].delta.content, end="")

Using openai:

from openai import OpenAI

client = OpenAI(
	base_url="https://api-inference.huggingface.co/v1/",
	api_key="hf_***"
)

messages = [
	{
		"role": "user",
		"content": "What is the capital of France?"
	}
]

stream = client.chat.completions.create(
    model="google/gemma-2-2b-it", 
	messages=messages, 
	max_tokens=500,
	stream=True
)

for chunk in stream:
    print(chunk.choices[0].delta.content, end="")

To use the Python client, see huggingface_hub’s package reference.

Code snippet example for conversational VLMs

Python
JavaScript
cURL

Using huggingface_hub:

from huggingface_hub import InferenceClient

client = InferenceClient(api_key="hf_***")

messages = [
	{
		"role": "user",
		"content": [
			{
				"type": "text",
				"text": "Describe this image in one sentence."
			},
			{
				"type": "image_url",
				"image_url": {
					"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
				}
			}
		]
	}
]

stream = client.chat.completions.create(
    model="meta-llama/Llama-3.2-11B-Vision-Instruct", 
	messages=messages, 
	max_tokens=500,
	stream=True
)

for chunk in stream:
    print(chunk.choices[0].delta.content, end="")

Using openai:

from openai import OpenAI

client = OpenAI(
	base_url="https://api-inference.huggingface.co/v1/",
	api_key="hf_***"
)

messages = [
	{
		"role": "user",
		"content": [
			{
				"type": "text",
				"text": "Describe this image in one sentence."
			},
			{
				"type": "image_url",
				"image_url": {
					"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
				}
			}
		]
	}
]

stream = client.chat.completions.create(
    model="meta-llama/Llama-3.2-11B-Vision-Instruct", 
	messages=messages, 
	max_tokens=500,
	stream=True
)

for chunk in stream:
    print(chunk.choices[0].delta.content, end="")

To use the Python client, see huggingface_hub’s package reference.

API specification

Request

Payload
frequency_penalty number Number between -2.0 and 2.0. Positive values penalize new tokens based on their existing frequency in the text so far, decreasing the model’s likelihood to repeat the same line verbatim.
logprobs boolean Whether to return log probabilities of the output tokens or not. If true, returns the log probabilities of each output token returned in the content of message.
max_tokens integer The maximum number of tokens that can be generated in the chat completion.
messages* object[] A list of messages comprising the conversation so far.
        content* unknown One of the following:
                 (#1) string
                 (#2) object[]
                         (#1) object
                                text* string
                                type* enum Possible values: text.
                         (#2) object
                                image_url* object
                                        url* string
                                type* enum Possible values: image_url.
        name string
        role* string
presence_penalty number Number between -2.0 and 2.0. Positive values penalize new tokens based on whether they appear in the text so far, increasing the model’s likelihood to talk about new topics
response_format unknown One of the following:
         (#1) object
                type* enum Possible values: json.
                value* unknown A string that represents a JSON Schema. JSON Schema is a declarative language that allows to annotate JSON documents with types and descriptions.
         (#2) object
                type* enum Possible values: regex.
                value* string
seed integer
stop string[] Up to 4 sequences where the API will stop generating further tokens.
stream boolean
stream_options object
        include_usage* boolean If set, an additional chunk will be streamed before the data: [DONE] message. The usage field on this chunk shows the token usage statistics for the entire request, and the choices field will always be an empty array. All other chunks will also include a usage field, but with a null value.
temperature number What sampling temperature to use, between 0 and 2. Higher values like 0.8 will make the output more random, while lower values like 0.2 will make it more focused and deterministic. We generally recommend altering this or top_p but not both.
tool_choice unknown One of the following:
         (#1) enum Possible values: auto.
         (#2) enum Possible values: none.
         (#3) enum Possible values: required.
         (#4) object
                function* object
                        name* string
tool_prompt string A prompt to be appended before the tools
tools object[] A list of tools the model may call. Currently, only functions are supported as a tool. Use this to provide a list of functions the model may generate JSON inputs for.
        function* object
                arguments* unknown
                description string
                name* string
        type* string
top_logprobs integer An integer between 0 and 5 specifying the number of most likely tokens to return at each token position, each with an associated log probability. logprobs must be set to true if this parameter is used.
top_p number An alternative to sampling with temperature, called nucleus sampling, where the model considers the results of the tokens with top_p probability mass. So 0.1 means only the tokens comprising the top 10% probability mass are considered.

Some options can be configured by passing headers to the Inference API. Here are the available headers:

Headers
authorization string Authentication header in the form 'Bearer: hf_****' when hf_**** is a personal user access token with Inference API permission. You can generate one from your settings page.
x-use-cache boolean, default to true There is a cache layer on the inference API to speed up requests we have already seen. Most models can use those results as they are deterministic (meaning the outputs will be the same anyway). However, if you use a nondeterministic model, you can set this parameter to prevent the caching mechanism from being used, resulting in a real new query. Read more about caching here.
x-wait-for-model boolean, default to false If the model is not ready, wait for it instead of receiving 503. It limits the number of requests required to get your inference done. It is advised to only set this flag to true after receiving a 503 error, as it will limit hanging in your application to known places. Read more about model availability here.

For more information about Inference API headers, check out the parameters guide.

Response

Output type depends on the stream input parameter. If stream is false (default), the response will be a JSON object with the following fields:

Body
choices object[]
        finish_reason string
        index integer
        logprobs object
                content object[]
                        logprob number
                        token string
                        top_logprobs object[]
                                logprob number
                                token string
        message unknown One of the following:
                 (#1) object
                        content string
                        role string
                 (#2) object
                        role string
                        tool_calls object[]
                                function object
                                        arguments unknown
                                        description string
                                        name string
                                id string
                                type string
created integer
id string
model string
system_fingerprint string
usage object
        completion_tokens integer
        prompt_tokens integer
        total_tokens integer

If stream is true, generated tokens are returned as a stream, using Server-Sent Events (SSE). For more information about streaming, check out this guide.

Body
choices object[]
        delta unknown One of the following:
                 (#1) object
                        content string
                        role string
                 (#2) object
                        role string
                        tool_calls object
                                function object
                                        arguments string
                                        name string
                                id string
                                index integer
                                type string
        finish_reason string
        index integer
        logprobs object
                content object[]
                        logprob number
                        token string
                        top_logprobs object[]
                                logprob number
                                token string
created integer
id string
model string
system_fingerprint string
usage object
        completion_tokens integer
        prompt_tokens integer
        total_tokens integer
< > Update on GitHub