AI should be shared.

We are pooling idle graphics cards to train open language models together. Turkish comes first. The result goes out under an open licence for everyone to use.

Volunteer How it works

We are at the research stage. No released model yet.

In AI, no one should hold this alone:

the code.

We will open the training code and the volunteer client under Apache-2.0.

the weights.

Model weights will be released so anyone can download and run them on their own machine.

the data list.

We will state plainly what we train on and the licence of every source.

the decisions.

Contributions will be on record; we will set the direction together.

We train it together. It belongs to everyone.

We will release code and model weights under Apache-2.0. We will list training-data sources and their licences openly.

Licence

Each square illustrates one volunteer graphics card. The real network does not exist yet.

Your graphics card is enough. No server room needed.

We freeze the base model and train only a small LoRA adapter. In our experiment peak memory was 5.2 GB; it fit on an 8 GB laptop card.

How it works

base ≈ 1.2 GB

adapter ≈ 40 MB

Areas are to scale (≈30:1)

We move less data. We get the same result.

Volunteers synchronise once every 25 steps, not every step. In our simulation about 25 times less data moved; the difference was 0.09% in bits per byte.

Our measurements

Synchronise every step: 25 syncs

Sparse sync: 1 sync

A 25-step window (illustration)

We write down our limits too.

Turkish language modelling improved, but we saw no improvement on TurkishMMLU. All small experiments: single run, single seed, no real network.

Roadmap

Base

35,6%

After LoRA

34,7%

TurkishMMLU 5-shot · dashed line: chance 20% · measurement noise ≈ ±1.6 points

What we have measured so far

We have no customer logos. We have measurements.

~25×less data moved

Sparse synchronisation on an adapter, 4 simulated workers. Simulation.

0.09%difference

Between sparse and every-step synchronised training, in bits per byte: 1.2166 vs 1.2155.

−4.3%Turkish bits per byte

20 minutes of LoRA on Qwen3-0.6B, on an 8 GB consumer graphics card.

0benchmark improvement

TurkishMMLU 35.6% → 34.7%, within noise. We cannot yet say “the model improved”.

Small experiments: single run, single seed, no real network. All tables on the Measurements page.

Where we are

00 · DONEResearch

Sparse-sync simulation on a small model.

01 · FIRST EXPERIMENTFirst language: Turkish

Short continued pre-training with LoRA on 8 GB.

02 · NEXTData and evaluation

More Turkish data, benchmarks.

03Volunteer client

Open source, signed updates.

04Pilot and open network

Real-network trials, then a first release.

No date commitments.

Why the Halay Ring?

In a halay everyone links arms with their neighbour; the circle has no beginning and no end. Our network will work the same way.

Every volunteer carries a lantern: the effort of your graphics card, a light. The light passes from hand to hand and no single hand holds it.

The circle: No leader; everyone stands in an equal place.
The hands: Contributions are linked and verify one another.
The lanterns: Each volunteer's light flows into the shared model.

Help we are looking for

This is not a one-person job. Right now we need most:

01Language benchmarksMore and better tests (Turkish first) to measure whether the model really improves. 02Data and licensingFinding and documenting open, clean, clearly licensed data sources (Turkish first). 03Security and verificationDesigning methods that catch faulty or malicious contributions. 04Volunteer clientWriting an open-source, signed client that runs safely on a computer.

Go deeper

How it worksBase, adapter and sparse synchronisation. RoadmapStages and known risks. FAQVolunteering, security, licence, data.

Be the first to know when the volunteer client is ready.

We use your email address only for Modelen announcements.

Volunteer

AI should be shared.AI should be sharedopenoursyours .