GPU capacity, without the data centre
SilentSwarm runs open language models across ordinary machines - your own, your partners', or units that heat buildings. You get an OpenAI-compatible API, and you know which machines did the work.
Building heating equipment? SilentSwarm for compute heat pumps
Why run it this way
Most teams rent GPUs by the hour in someone else's data centre. That works, until the model is too big for one card, the data must not leave the country, or the bill for idle capacity arrives.
- Bigger models on smaller cards. A model is split into stages across machines, so a pool of 24 GB cards can serve and fine-tune a model none of them could hold alone.
- You decide where it runs. Run the coordinator yourself and keep jobs on the machines you chose - in your office, at a partner, in the EU. Nothing goes to a third-party cloud unless you put a machine there.
- Heat that gets used. A pool can include compute heaters, machines whose waste heat warms a building. The energy every machine draws is metered, so the figure has a source.
What you can do with it today
The same platform members of the open beta use every day.
Chat and API
An OpenAI-compatible chat completions endpoint. Existing clients and SDKs work by changing the base URL and the key.
Fine-tuning across machines
LoRA fine-tuning split across the pool's GPUs, with checkpoints along the way and an exported model at the end, ready to serve.
Bring your own models
Open-weight models from Hugging Face. When a base model would not fit the pool for training, the coordinator can quantize it.
Private pools
Only members see a pool's machines and models. API tokens carry read, operate or admin scopes, and every change lands in the audit log.
Mixed hardware
Mismatched GPUs are welcome: each machine gets a share of the model weighted by its memory, and a health grade tells you which ones hold the pool back.
Agents can operate it
An MCP server gives AI agents the same actions as the web interface, with the caller's own permissions and nothing more.
Live from the SilentSwarm pool
Read from the coordinator that serves this page, not typed into it. Machine numbers are withheld while the pool is small, so they never describe one person's computer.
What exists, and what does not yet
We publish throughput only from real multi-machine runs, with the hardware and the links stated. The benchmark page has the numbers, the method and the limits.
Available today
- A self-hosted coordinator with pools, organizations, scoped API tokens and an audit log.
- Serving through an OpenAI-compatible API, across one machine or several.
- Fine-tuning split across machines, with checkpoints and exported models.
- Energy metering per machine, from the GPUs' own power readings.
- Published benchmarks from real runs across machines: Qwen3-14B on two home PCs, 11.6 tokens per second for eight users.
Not yet
- Usage billing and invoices.
- Service-level guarantees. It is an open beta.
- A managed capacity offer. Tell us what you need, and we will tell you honestly whether a pilot fits.
Questions teams ask
- Do our existing OpenAI clients work?
- Yes. The chat completions endpoint is OpenAI-compatible, so existing clients and SDKs work by changing the base URL and the key.
- Can it run models larger than our GPUs?
- Yes. A model is split into stages across machines, each holding a share weighted by its memory, so the cards together hold a model none of them could hold alone.
- Where does our data go?
- To the machines in the pool, and nowhere else. Run the coordinator yourself and nothing goes to a third-party cloud unless you put a machine there.
- How fast is it?
- We publish throughput from real runs across machines only: Qwen3-14B on two home PCs with older GPUs serves 11.6 tokens per second in total to eight users. The benchmark page has the method and the limits.
- Is there a service-level agreement or billing?
- Not yet. SilentSwarm is an open beta, without service-level guarantees or invoices. Tell us what you need, and we will tell you honestly whether a pilot fits.
- Can we use it commercially?
- Yes, with a commercial licence from NexPatch AI. The source is available under the PolyForm Noncommercial licence, which covers noncommercial use only.
Need GPU capacity? Tell us about it.
Which models, how much traffic or training, and where the data has to stay. We answer every message ourselves.
We use your details only to answer you - see the privacy policy. SilentSwarm is source-available under the PolyForm Noncommercial licence; commercial use needs a licence from NexPatch AI, which you can ask about in the same message.