Some pages are available in Turkish only for now.
Modelen documentation#
Modelen trains open language models with volunteer GPUs. The current method is adapter-DiLoCo: each volunteer trains a small LoRA adapter (r=16, ~10M parameters, ~40 MB) on top of Qwen3-0.6B-Base for a few steps; a coordinator merges the results and publishes the next round. Only the small adapter crosses the network.
This site is for people who want to run a coordinator or plug their own training system into Modelen.
Where to start#
| I want to… | Read |
|---|---|
| Try everything locally in 10 minutes | Quick start |
| Join the public directory as a coordinator | Becoming a coordinator |
| Know the exact wire contract | Protocol v1 and adapter format |
| Check my own implementation | Conformance test |
| Automate the portal (profile, heartbeat) | API introduction and API reference |
| Keep a coordinator healthy and safe | Heartbeat, Security |
The pieces#
- Volunteer app (Rust desktop app): downloads a task, trains locally, uploads the adapter.
- Coordinator: any HTTP service that implements protocol v1. A reference implementation and a stdlib-only minimal one are provided.
- Portal (modelen.tr): accounts, coordinator applications, the public directory and the Portal API.
Not
The Turkish documentation at the site root is the primary version; this English version covers the core pages. Pages not yet translated are linked from the Turkish site.