Fireworks AI Python SDK

0.19.20 · active · verified Sat Feb 28

Python SDK and OpenAI-compatible API for running open-source and proprietary LLMs on Fireworks AI infrastructure.

Warnings

breaking v1.0.0 alpha SDK (pip install fireworks-ai==1.0.0a*) is NOT stable. It has breaking changes vs v0.x and is still in pre-release as of Jan 2026. pip install fireworks-ai installs v0.19.20 (stable).
Fix: Use v0.19.20 for production. Pin explicitly: pip install fireworks-ai==0.19.20
breaking Model IDs use full path format: accounts/fireworks/models/<model-name>. Short names like 'llama-3-8b' will fail with 404.
Fix: Always use full model path. Find IDs at fireworks.ai/models
gotcha When using OpenAI SDK compatibility, set FIREWORKS_API_KEY env var but pass it explicitly as api_key= — the OpenAI SDK reads OPENAI_API_KEY by default and will silently use the wrong key.
Fix: api_key=os.environ['FIREWORKS_API_KEY'] explicitly in OpenAI() constructor
gotcha usage field is not exposed in OpenAI SDK streaming responses against Fireworks. Cast chunk to Any in TypeScript; in Python, usage will be None on stream chunks.
Fix: Use non-streaming for usage tracking, or read usage from the final chunk only
gotcha LiteLLM uses FIREWORKS_AI_API_KEY (with _AI_). Native SDK and OpenAI-compat use FIREWORKS_API_KEY. Different env var depending on integration.
Fix: Check integration docs. Set both if mixing LiteLLM with native/OpenAI SDK.
deprecated Build SDK (native fireworks.client) is marked deprecated in official docs. Fireworks now recommends OpenAI-compatible approach for new projects.
Fix: Migrate to OpenAI SDK with base_url override, or wait for stable v1.0.0

Install

pip install fireworks-ai Stable (recommended)
pip install fireworks-ai==1.0.0a20 v1 alpha (pre-release, breaking changes)
pip install openai OpenAI-compatible alternative (no native SDK needed)

Imports

Fireworks
```
from fireworks.client import Fireworks
```
v0.x stable import path

OpenAI (Fireworks compat)

from openai import OpenAI
client = OpenAI(base_url='https://api.fireworks.ai/inference/v1', api_key=os.environ['FIREWORKS_API_KEY'])

Must override base_url and api_key — defaults point to OpenAI

Quickstart

Chat completion using OpenAI-compatible interface (recommended for portability)

from openai import OpenAI

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

response = client.chat.completions.create(
    model='accounts/fireworks/models/llama-v3p1-8b-instruct',
    messages=[{'role': 'user', 'content': 'Hello'}]
)
print(response.choices[0].message.content)

view raw JSON →