pi-fireworks

A pi extension that registers Fireworks.ai models as a custom provider. All 200+ serverless models are available via the OpenAI-compatible API.

Features

• 200+ models — DeepSeek, Qwen, Llama, Mistral, GLM, Kimi, MiniMax, Gemma, and more.
• Vision models — Models with image input support (Qwen VL, InternVL, Llama Vision, etc.) are properly flagged.
• Reasoning models — DeepSeek R1, QWQ, and thinking models marked with reasoning: true.
• Zero config — Just set your API key and go.

Setup

1. Get your API key from https://fireworks.ai/account/api-keys
2. Set the environment variable:

   export FIREWORKS_API_KEY=fw_...

3. Install the package (see below)

Install

Add to your pi settings (~/.pi/agent/settings.json or .pi/agent/settings.json):

{
  "packages": [
    "https://github.com/Huijiro/pi-fireworks"
  ]
}

Or for local development:

{
  "packages": [
    "/path/to/pi-fireworks"
  ]
}

Then use /model to switch to any Fireworks model.

Models

Models are defined in src/models.json and loaded at startup. The list was scraped from the Fireworks model library and includes context window sizes, max output tokens, and input type support.

To update the model list, edit src/models.json directly.

Model families

Family	Count	Notable models
Qwen	60+	Qwen3.5 397B, Qwen3 Coder 480B, Qwen3 235B, Qwen VL
DeepSeek	25	DeepSeek V3.2, DeepSeek R1, R1 distills
Llama	25	Llama 4 Maverick/Scout, Llama 3.3, Llama 3.2 Vision
Mistral	17	Mistral Large 3 675B, Ministral 3, Mixtral
GLM	7	GLM-5, GLM-4.7
Gemma	7	Gemma 3 (with vision)
Kimi	4	Kimi K2.5, K2 Instruct/Thinking
MiniMax	4	MiniMax M2.5, M2.1
Others	60+	NVIDIA Nemotron, Cogito, InternVL, GPT-OSS, Phi, etc.

Development

# Install dependencies
npm install

# Type-check and lint
npm run check

# Auto-format
npm run format

License

MIT