Many computers.
One shared experiment.
Computers anywhere train separate blocks of one model and run it together. Join with one command, or try it below.
This is the small byte-level pilot, not the 512B model.
Bytes trained on
—Measuring the network rate…
Only bytes from accepted updates count, and text can repeat.
Network training throughput
Measuring…Bytes of text trained on per second, from accepted updates
Collecting at least 30 seconds of coordinator counter samples.
Anonymous contributions are tested before merging, including learning gain and held-out text checks. Trusted workers can bypass these checks. Rejected updates are not counted.
These are bytes, not the subword tokens larger models count.
Network inference throughput
Connecting…Bytes generated per second on the network, over the last hour's generations. Each byte travels through the blocks as a chain of hops, each run by a worker holding that block.
Work across the blocks
Updates merged per block. Blocks with the most left to learn get the most work.
Waiting for block metrics…
Evaluation history
Held-out denoising loss for the top block, in bits per byte. Lower is better.
No evaluation loaded yet.
Latest saved model sample
What the model wrote at the last evaluation, unedited.
Workers on this run
Waiting for worker records…
One card per worker. Green means it checked in within the last 45 seconds or merged an update in the last 2 minutes; amber means neither. Anyone leaving for good can take their own cards off the list.
Checking for an existing private owner session…
Signed in as the owner, your workers show in blue.
Sign in with the existing owner loginNo names, hostnames or addresses are shown. Hardware is what each worker reports about itself. Speed is measured by the coordinator.
Try a short continuation
Choose where inference runs: Automatic measures both real routes and selects the faster one; Single coordinator keeps computation on the coordinator; Distributed only requires participant workers and never substitutes the coordinator; Network keeps the existing coordinator-fallback behaviour. Every response stays pinned to one checkpoint snapshot. The sampler separately chooses the full chain or only the top block. It's a small model early in training, so expect broken text.
Do not enter private information or secrets. Your text goes to the coordinator and, whenever a distributed route is used, to participants' computers. Automatic calibration itself uses only fixed public text.
Your continuation will appear here.
Automatic measures one byte on single and distributed for each new snapshot/sampler and uses the faster route; the stream reports the selected route and probe timings.
One generation at a time across the whole network, so you may need to wait a moment. No account needed.
Join with one command
No account and nobody to ask. Your worker gets its own credential, trains one block at a time and serves that block for network inference. A GPU is optional. Contributions use compute, power and bandwidth.
- On Linux or macOS, run the command. It installs what it needs in
~/dbmoe-workerand starts. - On Windows, or to set it up by hand, use the second set of commands.
- Leave it running as long as you like. Stop it with Ctrl+C and start it again whenever you want; it keeps its credential.
- Leaving for good? Run the leave command: your workers come off the list and your credential ends. What you already contributed stays in the model. You can rejoin at any time.
Anonymous contributions are tested before they merge; trusted workers can bypass these checks. Your worker's card shows its CPU, RAM and GPU, and --no-share-hardware keeps them off the list. Keep credentials out of GitHub, screenshots and public messages.
Worker source and setup guide on GitHub
# Linux or macOS
curl -fsSL https://codex.opentransformers.online/dbmoe/join.sh | sh
# Windows, or by hand (Windows: py instead of python3, curl.exe in PowerShell)
python3 -m pip install torch --index-url https://download.pytorch.org/whl/cpu
curl -fsSL https://codex.opentransformers.online/dbmoe/code -o dbmoe.py
python3 dbmoe.py worker
# Options: --device cuda|rocm|xpu|mps --threads 4 --hours 1 --no-serve
# Leaving for good: takes you off the worker list
curl -fsSL https://codex.opentransformers.online/dbmoe/leave.sh | shUse only computers you own or are authorized to use. Source and setup guide: GitHub repository.
Use one identity across several machines
After installing the worker, create one contributor token. Copy its file securely to the other computers you own or are authorised to use. Start each machine with a different worker name; their contributions share the same identity.
# Create the shared identity once
python3 dbmoe.py enrol --token-file contributor.token
# On each machine, use its own name
python3 dbmoe.py worker --token @contributor.token --name machine-01Keep contributor.token private. Change machine-01 for each computer. To join anonymously with a separate identity, use the normal worker command above; no account is needed.