Spomin - Live KV cache compaction (Experimental for Qwen)

Spomin is an experimental router designed for Qwen models that aims to improve long-running sessions by replacing context with summaries directly in the KV cache. This method seeks to avoid frequent full compaction and reprocessing of context. In a Tetris benchmark, Spomin's managed approach nearly matched unmanaged performance, completing checks in 69.8 minutes versus 149.6 minutes, while generating 22.6% fewer tokens. The developer is seeking suggestions for repeatable benchmarks to measure quality, time, and token usage.