How it works
Not from scratch. From an open base.
Instead of training a model from the ground up, we train a small part that learns a new language (Turkish first) and split that part across volunteer computers.
We start from an open base
We take a suitably licensed model; in our first experiment Qwen3-0.6B-Base (Apache-2.0). Training from scratch takes months on one graphics card and does not fit in 8 GB of memory.
We train only a small adapter
With LoRA the base stays frozen. In our experiment the adapter was ≈ 40 MB and peak memory 5.2 GB. That fits on a laptop graphics card.
We merge updates sparsely
Each volunteer works on its own data slice for a while; only adapter differences are collected (DiLoCo). In our simulation about 25 times less data moved.
We verify contributions
Faulty or malicious updates can arrive. Our plan: verify the same work on several machines, check update size and consistency, and add small known test tasks. These mechanisms are still under development.
What happens on your computer
- The client will be open source; you can read what it runs.
- Updates will be signed.
- The client runs only the training job and does not access your files or personal data. This is a design principle; it must be verifiable once the client is released.
- You can stop at any time.
What we don't know
We have not measured how results change at 1 billion parameters and above, over real internet conditions, with different data slices, and with volunteer dropout and malicious contributions. Our measurements so far are on the Measurements page.