SOTA.app is one desk for the whole stack. Ask for an outcome — transcribe, infer, quantize, recall, research, imagine — and it sends the job to the instrument that is genuinely best right now: quality at speed, on your silicon, nothing off the device.
Pick a capability. The router scores every engine by quality / (1 + latency_s) and returns the SOTA pick — the same formula the Swift core ships.
Every piece is its own sovereign build. SOTA.app is the seam that makes them one machine.
| piece | capability | precision | latency | quality | SOTA |
|---|
A 1.58-bit CPU engine wins where it must be tiny; a 4-bit Metal kernel wins where it must be good. The right answer depends on the job — so the router is the product.
Every score is a number you can check: precision, median latency, quality. “SOTA” is a claim the app can be held to.
Cloud engines may exist in the pool, but the on-device pick wins by design. Your data ends at the machine.