Lumeo runs a single open model that fits a 16 GB card — with three personas: Solis to talk, Prism to code, Beacon to research. It reasons out loud, and you can watch it think.
100,000 tokens free every week · no credit card
The personas
The same weights under a different system prompt — each set up for a different job, and named for the way light behaves. Switch anytime; there’s nothing new to load.
Warm, direct, and quick — for questions, writing, and thinking out loud. Sees images, takes voice, searches the web, and calls your tools.
The same model tuned to write, refactor, and review code — complete, runnable answers that match your style, with tools and the web on hand.
The same model set to look things up — it reaches for web search, page fetch, and your MCP tools, and cites what it used.
| Detail | Value |
|---|---|
| Model | GPT-OSS-20B (open weights) |
| Parameters | 20B total · 3.6B active per token (MoE) |
| Serving precision | 4-bit MXFP4 |
| VRAM to serve | ~14 GB — fits a 16 GB card |
| Reasoning | Native — low, medium, high (streamable) |
| Vision | Images routed to a vision model |
| Voice | Whisper for audio input |
| Tools | MCP tool calling + web search |
| API | OpenAI-compatible /v1/chat/completions |
Quantized to 4-bit so the whole model serves comfortably on a single 16 GB card.
How we build
Lumeo runs GPT-OSS-20B — a 20B mixture-of-experts with only 3.6B active per token, served in 4-bit inside about 14 GB of VRAM. It fits a single 16 GB card and decodes like a much smaller model, and it reasons natively — so you can stream its thinking and hide it when you don't need it.
Around that core, requests route by what you send: images go to a vision model, audio through Whisper, and anything needing the outside world reaches for web search or a tool you've connected over MCP.
Requests route by modality — 4-bit, ~14 GB to serve
Honesty first
Lumeo is built on open weights (GPT-OSS) and still improving. It can be wrong, and it will sometimes make things up — please don't rely on it for anything important without checking. Your chats are tied to your account and used only to run the service — see our Privacy Policy and Terms of Service.
Create a free account and start a conversation in seconds.