Skip to content
superfluidlabs
Projects02The labProject Barter↗Engine Room↗
FIELD NOTES / SUPERFLUID LABS

Ideas worth
staying with.

Progress, experiments and occasional updates from Barter, Engine Room and whatever comes next.

Occasional emails. Confirm your address to join. Unsubscribe anytime.

How we use your email

Your address is used for Superfluid Labs updates and managed by Buttondown. Cloudflare provides hosting and spam verification; your IP address is used for signup abuse checks. We do not sell your address. Use the unsubscribe link in any newsletter to leave. You can manage your subscription and request removal through Buttondown.

Buttondown privacy policy ↗
superfluidlabs

Independent thinking.
Ideas in motion.

Explore projects ↗
© 2026 Superfluid LabsResearch. Build. Repeat.Back to top ↑
Site measurement

When enabled, Cloudflare Web Analytics measures visits and page performance without cookies. We respect your browser’s Do Not Track and Global Privacy Control preferences.

← All projects

SUPERFLUID LABS / PROJECT 001

Project Barter.

Research prototype

Spare compute.
Shared possibility.

What if the computers we already have could work together to make large AI models more accessible?

Barter explores API access to large models using contributors’ spare compute.

Explore the idea ↓

Many machines. One request.

THE PROPOSED APPROACH
[ ]DeveloperOne APIContributor AContributor BContributor C

A request comes in

The proposed API would give a developer one place to send a model request, without managing contributor machines directly.

Conceptual architecture. Illustrative machine count; not a live system.

THE QUESTION

Can a model go beyond
the memory of one machine?

The first goal is to split a model across contributors’ machines and serve correct responses through an API. That premise is still being tested.

Where the work stands.

A development prototype, with the core distributed inference premise still to prove.

01 / REFERENCE

Establish the baseline.

A sufficient-memory reference test passed. A matched 2 GiB control failed on memory, showing the test’s memory budget binds.

Tested
02 / DISTRIBUTION

Prove the split.

A correct response from a model split across independent machines remains unproven. This is the next technical milestone.

In research
03 / ACCESS

Make it useful.

The intended audience includes developers, students, AI researchers and small teams. Reliability, economics and demand still need validation.

To validate

Research status: 5 November 2026. The tested model fits a sufficient-memory host; the control establishes a budget constraint, not a model too large for one machine. No public API or contributor enrollment is available.

A few things to know.

What is Project Barter?+

Project Barter is a research prototype from Superfluid Labs exploring API access to large AI models through contributors’ spare compute. Distributed inference correctness, reliability and economics are still being validated.

Can I use the API today?+

No. Barter is a development prototype. Public API access has not launched.

Can I contribute my computer?+

Contributor onboarding is not open. Hardware requirements, network requirements and contributor payments remain under development.

Is this the project previously called Warren?+

Yes. Project Barter is the new name for Warren, which was earlier called KILN. It is now a project under Superfluid Labs.

What is the intended business model?+

The intention is to charge for API usage, pay contributors and retain a share to support the project. Pricing and contributor rewards have not been finalized.

A physical connection to the idea.

Project Engine Room explores connected gyms where verified workouts could earn credits for AI use through Barter. Its energy recovery and rewards are separate flows; the integration is a concept, not an available service.

Explore Engine Room ↗
A PROJECT BY SUPERFLUID LABS

One question.
A whole new possibility.

Back to the lab ↗