Cuna: inference optimizations for open models.

What we built

Cuna applies inference optimizations to compatible open models. GLM-5.2 is the example shown in this test.

The recorded Cuna run produced about 260 output tokens per second.

Our inference optimizations sit around the model: request routing, batching, kernel execution, and memory movement.

They are not tied to GLM-5.2. They can be applied to other compatible open models.

The OpenRouter route in the recording produced about 58–60 output tokens per second. Cuna was roughly 4.4× faster in this run.

This is one observed test, not a universal OpenRouter benchmark. OpenRouter speed changes with provider, load, quantization, and routing.

Cuna Model API is planned for open-model inference.

It will bring these inference optimizations to an API: more tokens per second at a lower cost. Availability and pricing will be announced separately.