Models

Every model the network will run, with the measured numbers the planner uses to decide what fits a machine. Weight sizes are summed from the real files; commit shas are pinned; KV figures are per token at fp16.

Reading the memory column

Estimated memory = weights + KV cache at that context + activation peak + runtime overhead. For a model with sliding-window layers the windowed part stops growing at the window; for a recurrent hybrid the recurrent state is constant. These are feasibility estimates — the backend derives the final figure and the agent's plan --explain shows the arithmetic per machine.

Clients ask for a tier alias (deference-tiny … deference-large) or a model id; the scheduler routes to a machine serving it.

Variants marked mtplx run on Apple Silicon (M4 and newer) under MTPLX, which serves MLX weights with the model's own multi-token-prediction heads. Powered by MTPLX.