Fullmetal
A distributed network for running open-source LLMs, without renting a datacentre.
- Client
- Fullmetal.Ai
- Sector
- AI infrastructure
- Our role
- Platform and distributed agent network build
- Status
- Live, released as a research preview
The brief
What was breaking
New open-source language models arrive weekly, and running them yourself is the only way to keep your prompts private. That is also the expensive way. Fullmetal set out to make model choice and privacy affordable for people building personal projects, by pooling machines instead of buying them.
- The number of open-source models released each week makes choosing and testing one a project in itself.
- Self-hosting an LLM requires hardware that is idle most of the time, which is difficult to justify for a personal project.
- Existing AI APIs mean sending your prompts to a third party, which is exactly what people self-host to avoid.
What we built
The system we delivered
A distributed agent network behind one API
Requests are served by a pool of Fullmetal-hosted and community-contributed machines, each carrying at least 8GB of RAM or VRAM, so model availability comes from the network rather than from one operator's hardware budget.
Idle capacity that earns
Agent owners choose whether their machine serves public prompts while idle, and earn tokens for the prompts it handles.
Encrypted prompts and responses
Traffic is encrypted end to end so the privacy argument for self-hosting still holds when the compute belongs to someone else.
Agent operations in the open
Operators see each agent's status, creation date, prompts served, coins earned, and public-serving toggle, so participation in the network is legible rather than a black box.
In-browser inference with WebLLM
Models including Llama 2 and RedPajama-INCITE run as quantised builds directly in the browser, with cached parameters making subsequent visits fast.
Built with
The stack behind it
The lead tier is what makes this system what it is. Everything under it is the platform that carries it.
AI and models
Platform
Security
What changed
The result in production
Model choice without the hosting bill
Pooled community machines put a range of open-source models within reach of individual projects.
Privacy that survives the move off your own hardware
Encrypting prompts and responses removes the trade-off between not self-hosting and not sending your data to a third party.
An incentive to keep the network supplied
Token rewards for serving public prompts give idle high-spec machines a reason to stay in the pool.
The screens
What it looks like in use
Same problem?
Let's scope what this would look like for you
Start with a two-week Discovery Sprint. We map your highest-value workflows and deliver a prioritised pilot roadmap grounded in what we have already shipped.