Supercomputers

Training your own model used to take an ML team and a cluster. Now it takes a dataset, a benchmark and a budget. Your agent does the rest.

Watch your agent train

Loss, evals, logs and spend, live. Your agent reads the same numbers you do, and acts on them.

Training shouldn't need an infra team

Your agent sets up the run, watches it and keeps the best model. You set the budget.

• GPUs

  The latest GPUs, from a single card to a full cluster.

• Training

  Hand your agent a dataset and a budget. It fine-tunes, and never spends a cent more than you allowed.

• Jobs

  Fire off data prep, batch inference or a sweep, and let your agent watch it land.

• Volumes

  Super fast storage for datasets and checkpoints, shared by every run.

• Evals

  Score every version against your own benchmark, and keep the winner.

Designed and loved by agents

Plain commands, great docs and output your agent can read and act on.

# Push your dataset to super fast block storage
$ div storage put training-data ./data

# Fine-tune on 8 H100s with a $50 budget
$ div run --gpu h100 --gpus 8 train.py::run --max-spend 50

# Stream logs your agent can act on in realtime
$ div logs run-k4t9 -f

# Score the new version and see where it ranks
$ div eval run held-out --model llama-ft:v4 --gpu h100 eval.py
$ div eval report held-out

Pricing

Pay for what your agents use, billed by the minute.