Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.fireworks.ai/llms.txt

Use this file to discover all available pages before exploring further.

You can use the OpenAI Python client library to interact with Fireworks. This makes migration of existing applications already using OpenAI particularly easy. For Anthropic SDK support, see Anthropic compatibility.

​
Specify endpoint and API key

​
Using the OpenAI client

You can use the OpenAI client by initializing it with your Fireworks configuration: 
from openai import OpenAI

# Initialize with Fireworks parameters
client = OpenAI(
    base_url="https://api.fireworks.ai/inference/v1",
    api_key="<YOUR_FIREWORKS_API_KEY>",
)
You can also use environment variables with the client: 
import os
from openai import OpenAI

# Initialize using environment variables
client = OpenAI(
    base_url=os.environ.get("OPENAI_API_BASE", "https://api.fireworks.ai/inference/v1"),
    api_key=os.environ.get("OPENAI_API_KEY"),  # Set to your Fireworks API key
)

​
Using environment variables

export OPENAI_API_BASE="https://api.fireworks.ai/inference/v1"
export OPENAI_API_KEY="<YOUR_FIREWORKS_API_KEY>"

​
Alternative approach

import openai

# warning: it has a process-wide effect
openai.api_base = "https://api.fireworks.ai/inference/v1"
openai.api_key = "<YOUR_FIREWORKS_API_KEY>"

​
Usage

Use OpenAI’s SDK how you’d normally would. Just ensure that the model parameter refers to one of Fireworks models.

​
Completion

Simple completion API that doesn’t modify provided prompt in any way: 
from openai import OpenAI

client = OpenAI(
    base_url="https://api.fireworks.ai/inference/v1",
    api_key="<YOUR_FIREWORKS_API_KEY>",
)

completion = client.completions.create(
    model="accounts/fireworks/models/llama-v3p1-8b-instruct",
    prompt="The quick brown fox",
)
print(completion.choices[0].text)

​
Chat Completion

Works best for models fine-tuned for conversation (e.g. llama*-chat variants): 
from openai import OpenAI

client = OpenAI(
    base_url="https://api.fireworks.ai/inference/v1",
    api_key="<YOUR_FIREWORKS_API_KEY>",
)

chat_completion = client.chat.completions.create(
    model="accounts/fireworks/models/llama-v3p1-8b-instruct",
    messages=[
        {
            "role": "system",
            "content": "You are a helpful assistant.",
        },
        {
            "role": "user",
            "content": "Say this is a test",
        },
    ],
)
print(chat_completion.choices[0].message.content)

​
Fine-tuning compatibility

Fireworks fine-tuning uses the same OpenAI-compatible chat completion format for training data. If you have datasets formatted for OpenAI SFT, they work on Fireworks with no conversion required — the same messages array with role, content, tool_calls, and weight fields. Fireworks also supports additional features in the training schema:
  • Thinking traces via reasoning_content on assistant messages (for models like DeepSeek R1 and Qwen3)
  • Per-message weights to control which turns the model learns from
  • Per-sample weights for weighted training
  • Vision inputs using the same OpenAI-compatible multimodal content format
To get started with fine-tuning, see the Supervised Fine-Tuning guide or the Fine-Tuning overview.

​
API compatibility

​
Differences

The following options have minor differences:
  • max_tokens: behaves differently if the model context length is exceeded. If the length of prompt or messages plus max_tokens is higher than the model’s context window, max_tokens will be adjusted lower accordingly. OpenAI returns an invalid request error in this situation. Control this behavior with the context_length_exceeded_behavior parameter:
    • truncate (default): Automatically adjusts max_tokens to fit within the context window
    • error: Returns an error like OpenAI does

​
Token usage for streaming responses

OpenAI API returns usage stats (number of tokens in prompt and completion) for non-streaming responses but doesn’t for the streaming ones (see forum post). Fireworks API returns usage stats in both cases. For streaming responses, the usage field is returned in the very last chunk on the response (i.e. the one having finish_reason set). For example:
cURL
curl --request POST \
     --url https://api.fireworks.ai/inference/v1/completions \
     --header "accept: application/json" \
     --header "authorization: Bearer $API_KEY" \
     --header "content-type: application/json" \
     --data '{"model": "accounts/fireworks/models/starcoder-16b-w8a16", "prompt": "def say_hello_world():", "max_tokens": 100, "stream": true}'

data: {..., "choices":[{"text":"\n  print('Hello,","index":0,"finish_reason":null,"logprobs":null}],"usage":null}

data: {..., "choices":[{"text":" World!')\n\n\n","index":0,"finish_reason":null,"logprobs":null}],"usage":null}

data: {..., "choices":[{"text":"say_hello_","index":0,"finish_reason":null,"logprobs":null}],"usage":null}

data: {..., "choices":[{"text":"world()\n","index":0,"finish_reason":"stop","logprobs":null}],"usage":{"prompt_tokens":7,"total_tokens":24,"completion_tokens":17}}

data: [DONE]
Note, that if you’re using OpenAI SDK, they usage field won’t be listed in the SDK’s structure definition. But it can be accessed directly. For example:
for chunk in client.chat.completions.create(stream=True, ...):
    if chunk.usage:  # Available in final chunk
        print(f"Tokens: {chunk.usage.total_tokens}")