Fireworks AI Provider

Fireworks AI specializes in FAST open-model inference using their FireAttention kernel — often 3-5x faster than competitors.

Install

bashpip install "mtpx[fireworksai]"

This installs the openai SDK. Alternatively, install the native Fireworks SDK:

bashpip install fireworks-ai

API Key Setup

Option 1: .env file (recommended)

1. Install dotenv support:

bash   pip install python-dotenv

Or with the MTP extra:

bash   pip install "mtpx[dotenv]"

1. Create a .env file in your project root:

text   FIREWORKS_API_KEY=fw_your_key_here

1. Load it in your code before creating the provider:

python   from mtp import Agent

   Agent.load_dotenv_if_available()  # reads .env file

Get a free API key at fireworks.ai (free credits on signup).

Option 2: System environment variable

bash# Linux/macOS
export FIREWORKS_API_KEY="fw_..."

# Windows PowerShell
$env:FIREWORKS_API_KEY="fw_..."

Quick Start

pythonfrom mtp import Agent
from mtp.providers import FireworksAI

Agent.load_dotenv_if_available()  # loads FIREWORKS_API_KEY from .env

provider = FireworksAI(model="accounts/fireworks/models/llama-v3p3-70b-instruct")
tools = Agent.ToolRegistry()
agent = Agent(provider=provider, tools=tools)

reply = agent.run_loop("What is 25 * 4 + 10?")
print(reply)

Parameters

Capabilities

Recommended Models

• accounts/fireworks/models/llama-v3p3-70b-instruct — Best overall (default)
• accounts/fireworks/models/llama-v3p1-405b-instruct — Most capable
• accounts/fireworks/models/qwen2p5-72b-instruct — Top reasoning
• accounts/fireworks/models/deepseek-v3 — Best reasoning
• accounts/fireworks/models/mixtral-8x22b-instruct — Fast + smart
• accounts/fireworks/models/firefunction-v2 — Purpose-built for tool calling

Structured Output

Fireworks supports native JSON schema output:

pythonprovider = FireworksAI(
    model="accounts/fireworks/models/llama-v3p3-70b-instruct",
    response_format={
        "type": "json_schema",
        "json_schema": {
            "name": "result",
            "schema": {
                "type": "object",
                "properties": {
                    "answer": {"type": "number"},
                    "explanation": {"type": "string"},
                },
                "required": ["answer", "explanation"],
            },
        },
    },
)

Full Example

pythonfrom mtp import Agent
from mtp.providers import FireworksAI

Agent.load_dotenv_if_available()

provider = FireworksAI(
    model="accounts/fireworks/models/llama-v3p3-70b-instruct",
    temperature=0.0,
    max_tokens=4096,
)

tools = Agent.ToolRegistry()
agent = Agent(provider=provider, tools=tools, debug_mode=True)

reply = agent.run_loop(
    "Calculate (25 * 4) + 10",
    max_rounds=3,
)
print(reply)

Notes

• Fireworks prefers the native fireworks-ai SDK, falls back to OpenAI client at https://api.fireworks.ai/inference/v1.
• Model IDs must be account-qualified (e.g., accounts/fireworks/models/...).
• firefunction-v2 is purpose-built for function/tool calling and may give better results for agent workflows.
• When response_format is set, structured_output_support reports native_json_schema capability.

Source

src/mtp/providers/fireworks_provider.py