Get started
Compute

Workload

AI training and fine-tuning

GPU machines for training models and fine-tuning open ones.

Not available yet

What it is

Training builds a model's weights from data. Fine-tuning starts from an open model and adjusts it on your own examples, often with a small adapter such as LoRA.

Jobs run for hours or days on one or more GPUs. What matters is GPU memory, the link between GPUs and the price per GPU hour. Interruptible capacity costs less if the job saves checkpoints and can resume.

What we offer today

Nothing yet. Silicon Network does not run training or fine-tuning jobs.

If you train a LoRA adapter elsewhere, you can register it and serve it through the inference API.

Sources

The kinds of source that can run this workload, now or later. No provider is named until it is connected.

  • Silicon Network operators

    Independent machines that operators add to the network, today Macs running the Silicon Network engine

    Planned

  • Decentralized networks

    Open networks of independent machines

    Planned

  • GPU clouds and marketplaces

    Companies that rent out GPU machines

    Planned

  • Public clouds

    The large general-purpose clouds

    Planned

On operators' own Linux GPU machines, training jobs would run through the node agent, in containers with no network unless the operator allows it. The agent is built and switched off.

Status

Not available yet. Nothing is listed or sold until a source is connected and checked.

Open in the chooserTell us what you need

The chooser opens with AI training and fine-tuning selected.

Other workloads