Skip to content

Benchmarking

Measurements and explanations of configured systems, with methods and limits visible alongside the findings.

Research article · local inference

How Strata uses GPU, RAM and SSD →

Expert routing, three memory pools and KV cache, explained visually.

 Benchmark report · results and reproduction

Strata through a 256K window →

The configured RTX 4080 study: prefill, prefix reuse, decode and exact retrieval.