Does citizen data leave the city?

Not when the city runs it itself. The coordinator is a server of the city's, the machines in the pool are the city's own, and prompts, documents and trained models stay on them.

Which machines can join?

Machines running Linux with an NVIDIA GPU of 4 GB or more; mismatched cards are fine. Each joins with one command, without administrator rights, and connects out to the coordinator, so it works behind the city's firewall without opening a port. Windows is not supported yet.

Can a school's computers join only after hours?

Yes. Each machine can have weekly windows in which it is not available. Fifteen minutes before a window starts, the pool moves its work elsewhere at the next saved step and takes the machine out, and it comes back in when the window ends.

Is it free for a city administration?

SilentSwarm is source-available under the PolyForm Noncommercial licence, which names government institutions as permitted users, however the work is funded. A municipal utility that sells its services needs a commercial licence from NexPatch AI. Tell us about your case, and we will say which applies.

Can the heat warm our buildings?

A machine's heat warms the room it stands in. To put it into a building's heating, a compute heater - a GPU unit built into the heating system - is needed, and the platform meters each unit's energy and listens to each building's heat demand. See SilentSwarm for compute heat pumps.

We publish throughput from real runs across machines only: Qwen3-14B on two home PCs with four older GPUs serves 3.3 tokens per second to one user and 11.6 in total to eight. The benchmark page has the method and the limits.

For cities, municipalities and municipal utilities

AI for the city, on the city's own machines

SilentSwarm joins the GPUs a city already runs - in schools, libraries and offices - into one private cluster that serves and fine-tunes open language models. The data stays on the city's network, and the heat can warm the buildings it runs in.

Talk to us about a pilot Count what you already own

Inside the city networkno cloud contract needed

After hoursmachines join on a schedule

Meteredenergy and heat per machine

Building heating equipment? SilentSwarm for compute heat pumps

Why a city should run it this way

Cities want language models for their staff and their services, but not every request should leave for a cloud abroad, and new GPU servers are a budget line of their own. Meanwhile many of the city's computers stand idle every night.

The data stays in the city. The coordinator runs on a server of the city's, and the machines in the pool are the city's own. Prompts, documents and trained models never leave its network, and the source code can be audited.
Use what is already paid for. A school's computer room or an office's workstations join the pool after hours. Availability windows keep a machine out of the pool while people need it.
Heat where the city needs it. Every machine's GPU energy is metered, and a building's heat demand steers the work towards the machines whose heat is wanted - a school in winter, a swimming pool all year.

What a city can do with it today

The same platform members of the open beta use every day, on hardware the city chooses.

An assistant for staff

A chat interface for open language models: draft a letter, summarise a long document or translate a notice, without sending the text to an outside service.

An API for the city's own apps

An OpenAI-compatible endpoint, so citizen portals and internal tools built for a cloud model switch over by changing the base URL and the key.

Models that know the city's language

LoRA fine-tuning across the pool's GPUs on the city's own texts, with checkpoints along the way and an exported model at the end.

Machines on a schedule

Weekly availability windows per machine: a school's computer room works for the pool at night and at weekends and is out of it again before the first lesson.

Heat metering and demand

GPU energy metered per machine, and a heat-demand signal per building that sends work to the machines whose heat is wanted.

Separate pools per department

Private pools, organizations and scoped API tokens keep departments apart, and every change lands in the audit log.

How it fits together

Four pieces, and every one of them belongs to the city.

A coordinator in the city's data room. One small always-on server, which needs no GPU. It holds the accounts, the pools and the models, and decides where work runs.
Machines that connect out. Each machine joins with one command and connects out to the coordinator over a private mesh, so no port has to be opened into a school or an office.
Staff and apps that ask. People use the chat, and the city's software uses the API. A model too big for any one card is split across several, and the answer streams back.
Buildings that take the heat. Where a machine sits in a building that wants heat, its waste heat stays there, and the coordinator knows how much energy each machine turned into it.

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, and the heat figures count GPU power only - the whole machine gives off somewhat more.

··· machines online

··· GB of GPU memory

··· kWh of heat so far

··· GPU-hours of work

What the city already owns

Count the machines with a GPU in schools, libraries and offices, and the hours they stand idle. Drag a slider or type a value, and the result redraws.

GPU memory in all: -
Left for models: -
Largest model in 4-bit: -
GPU-hours a year: -
Heat a year: -

A rough estimate, not a plan. It counts serving a model in 4-bit with room for a conversation, keeps 1 GB of every card for its driver and assumes one GPU per machine. Fine-tuning needs more memory, and every extra machine in a model's chain makes each answer a little slower. The coordinator checks each real job against the machines it would run on, and the benchmark page shows what two PCs measured.

The address bar keeps the values you set, so a link to this page shares them.

What exists, and what does not yet

SilentSwarm is in open beta. A city pilot works best when both sides know where the edges are.

Available today

A self-hosted coordinator with pools, organizations, scoped API tokens and an audit log.
Availability windows per machine, and one-command onboarding behind any firewall.
Open models served through an OpenAI-compatible API and fine-tuned across machines.
Energy metering per machine, and the heat-demand signal that steers where work runs.

Not yet

Windows machines. A machine needs Linux and an NVIDIA GPU today; Windows is untested.
Search over the city's documents. There is no document connector yet: a model answers from what it was trained on and what a request contains.
Certifications such as BSI C5 or ISO 27001, and service-level guarantees. It is an open beta.
An offline install. A machine installs PyTorch from PyPI when it joins, and a coordinator without internet needs its models put on it by hand.

Questions cities ask

Does citizen data leave the city?: Not when the city runs it itself. The coordinator is a server of the city's, the machines in the pool are the city's own, and prompts, documents and trained models stay on them.
Which machines can join?: Machines running Linux with an NVIDIA GPU of 4 GB or more; mismatched cards are fine. Each joins with one command, without administrator rights, and connects out to the coordinator, so it works behind the city's firewall without opening a port. Windows is not supported yet.
Can a school's computers join only after hours?: Yes. Each machine can have weekly windows in which it is not available. Fifteen minutes before a window starts, the pool moves its work elsewhere at the next saved step and takes the machine out, and it comes back in when the window ends.
Is it free for a city administration?: SilentSwarm is source-available under the PolyForm Noncommercial licence, which names government institutions as permitted users, however the work is funded. A municipal utility that sells its services needs a commercial licence from NexPatch AI. Tell us about your case, and we will say which applies.
Can the heat warm our buildings?: A machine's heat warms the room it stands in. To put it into a building's heating, a compute heater - a GPU unit built into the heating system - is needed, and the platform meters each unit's energy and listens to each building's heat demand. See SilentSwarm for compute heat pumps.
How fast is it?: We publish throughput from real runs across machines only: Qwen3-14B on two home PCs with four older GPUs serves 3.3 tokens per second to one user and 11.6 in total to eight. The benchmark page has the method and the limits.

Planning AI for your city? Let's talk.

Tell us about your machines, what you want to use it for 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.