Kern

Kern packages machine learning models as verified GPU programs consisting of a manifest, kernels, and weights. It separates inference engine from model configuration, allowing agents to optimize performance per GPU and workload without modifying shared codebases. The system enables rapid optimization: agents have achieved 17% throughput gains on Qwen models, brought new model families online in 2.5 hours, and generated kernel sets across parallel agent runs. Kern is built for teams running inference at scale who need per-GPU tuning and model-specific optimization without the complexity of maintaining massive shared engines.