Most teams train in one of two ways: rent a cluster and staff an ML team, or accept the model the vendor shipped. Inference now fits on a laptop more often than training does. Caletta is building a third path for fine-tuning, distillation, and adaptation on operator-owned data.
A laptop will not carry every training run. The proposed bridge keeps data, policy, and resulting weights under operator control while borrowing CUDA capacity for the expensive steps. Local and remote compute are two parts of one run, not opposing deployment choices.
Ownership follows the loop
Ownership does not require every FLOP to run on local hardware. It requires control of the inputs, run policy, evaluation, and resulting weights. Borrowed compute should be replaceable; the capability produced by the run should remain portable.
What exists today
The Go-native MLX runtime performs inference and training on Apple silicon and has a Linux CUDA build path. mlx-go-iroh, built on go-iroh, supplies the public peer-to-peer transport. The self-hostable training coordinator serves a training API on the operator's machine and is publication pending. Membership, quorum, and signed-receipt code is built and tested.
What is not wired yet
The missing connection is the tensor data plane. Two Macs have exchanged and merged adapter weights across three measured rounds (Two Macs, one adapter). A Mac and a remote CUDA host have not yet completed a cooperative training run through this stack. CUDA is a build option, not a shipped and continuously tested binary.
The next acceptance test wires training tensors onto the existing transport and runs them end to end between a Mac and a physical CUDA host. Until that passes, this is a control plane with two compute targets, not a working bridge between them.