About HiveBear
Your laptop can’t run a 70B model alone. A few machines together can.
The idea
Most of the world’s compute sits idle. Meanwhile the models worth running keep outgrowing the machines most people own, and the usual answer is to rent someone else’s GPU and send your prompts to it.
HiveBear takes the other route. Devices pool spare capacity into a peer-to-peer mesh, a model’s layers are split across peers, and activations are passed between them. Every device that joins makes the network able to run something bigger.
How it works
HiveBear profiles your hardware and runs what fits on your own machine, with llama.cpp, Candle, ONNX or MLX.
When a model is too large, its layers are distributed across mesh peers and the work is shared.
Chat history never leaves your device. Contributing is optional, and you choose when it runs.
Open source
The client, the mesh, and the CLI are MIT-licensed and developed in the open. Reading the code is the best way to check that the privacy claims on this site are real.
BeckhamLabsLLC/HiveBearWho builds it
HiveBear is built by Beckham Labs LLC. It is early software: the local inference path is solid, and the distributed mesh is actively being built out. If something is broken, we would rather hear about it than have you work around it — open an issue.
