Welcome to Actual Computer
We are building the software infrastructure for the ever-present future: when people come across a device which can't utilize local AI, they will dismiss them as artifacts of an earlier age. Something that should probably be in a museum.
A device which can't think will not be seen as an Actual Computer.
We think preparing for this is not just realistic, but a problem worth taking seriously.
A lot of AI infrastructure still carries an old assumption: build for scale within the context of the cleanest system, and push software which more or less forces the hardware to look uniform. Sometimes that works! But a lot of the time, it just means you throw away useful compute, or leave it sitting idle, build around the ideal case, and then act surprised when the production systems and local AI don't mix. Most people have non-ideal setups.
Mass local AI adoption has yet to occur, and the thought process around developing inference software -- because of its relative novelty & smallness has the tendency to drive people to treat the entire scope as just "developer tools". We disagree with this. We believe that the structure supporting the future of machine intelligence should look like breaker boxes and transformers, less like apps.
What we are building
Actual Computer is building a system for inference that can coordinate real hardware as it exists, allowing for hosting GGUF models on local devices as well as across them, while striving to produce realtime tokens/second on everything from frontier-capable hardware to low-resource devices.
The process is simple: sign up, get a one-line command to install Actual on your machine of choice (Mac/Linux/PC), install some models via the interface, and get functional OpenResponses / "big provider"-compatible endpoints which we route for you without ever seeing or storing your inference requests.
Fully private, functional, always-on compute at home -- anywhere you want it.
Fire up a home cluster with a powerful model and access it from anywhere.
Local, global, whichever way you prefer. After all, it's your hardware.
We want distributed inference to work across heterogeneous machines without requiring people to flatten everything into a fake homogenous cluster first. Different devices have different strengths. Different nodes have different failure modes. Different workloads want different partitioning strategies. We think a good system should be able to see that clearly and make useful decisions instead of reaching for the same blunt abstraction every time.
To accomplish this, we are developing our very own inference engine from the ground-up to accomplish record-setting performance across a variety of different device configurations.
To enhance this, we have many surprises in store -- from rewarding ways for developers to contribute directly to our progress, to allocated bonuses for dedicating your devices to a consumable compute pool which we share with upstream providers.
Our mission & values
Our mission is to make serious inference infrastructure that can use the hardware people actually have.
Not ideal hardware. Not fantasy hardware. Actual hardware.
We want researchers, engineers, and small teams to be able to compose useful inference systems out of mixed fleets and still get something that feels disciplined, intelligible, and fast. We want the software layer to meet reality where it is and make better use of it.
And we want that software to work well at massive scale as well.
For satisfying any request while allowing people to access their own, local, private, encrypted inference.
That is the bet behind Actual Computer.
Join our closed beta today by providing your email in the sign-up link on our homepage: https://actual.inc/.