Roadmap

We write the stages, not the dates.

When a stage ends we will publish what we measured together with what we could not find.

00 · DONEResearch

Sparse synchronisation on a small model and an adapter-scale DiLoCo simulation.

01 · FIRST EXPERIMENTFirst language: Turkish

Short continued pre-training with LoRA on Qwen3-0.6B-Base on 8 GB.

02 · NEXTData and evaluation

More Turkish data, English replay mix, tokenizer, a 1.7B base.

03Volunteer client

Open source, signed updates, contribution verification; alpha.

04Closed pilot

Training over a real network with a small group of volunteers.

05Open network and first release

Weights, code and list of data sources under Apache-2.0 / applicable licences.

Known risks

Research suggests weights trained with sparse synchronisation may not combine well with later standard training. We will test this ourselves.
Training-data licences (e.g. CC BY-SA data) need legal review for the weights we release. We will announce the outcome on the Licence page.
Open models already exist for many languages. Our difference has to be measurable improvement and community ownership.