Archipelago
Our own models, an archipelago of binary spiking-neural islands evolved to mimic the frontier models, then surpass them. An island surfaces here, public and pass-free, the moment its measured accuracy crosses 95%. No accuracy claim is made until it is earned.
frontier inference
A GPU cluster. Racks of accelerators, megawatts, a metered API, and your prompt running on someone else's silicon.
semurg, CPU-only
2.84 trillion synaptic ops/s. Binary spiking neural nets, fully recurrent (state carried across steps, no history reload, O(1) per step), our own models, on the CPU you already own.
Trillions of synaptic operations per second, no GPU, no cloud, your data never leaves.
SAOPS measured live via VPOPCNTQ on this box. An island graduates to public at 95% measured accuracy, earned, never claimed.
No mimic has crossed 95% yet, so nothing is published here. That's the honesty bar: we don't claim a model is ready until it measurably is. Bring-up status (real numbers, in training) is in the gated lab.