Superlinked has open-sourced an inference cluster architecture designed for small language models that replaces top-down routing bottlenecks with worker-driven pull queuing, doubling GPU throughput and drastically cutting latency on accessible hardware.
FireBrief
FireBrief