Cacheon

About Cacheon

Cacheon is a live competition platform designed to accelerate large language model (LLM) inference by incentivizing developers (miners) to submit optimized GPU kernels. Participants write Triton or Cutedsl kernels targeting specific operations or blocks within a fixed model, aiming to outperform the baseline sglang implementation in terms of throughput while maintaining output fidelity and accuracy. Successful submissions are validated through a rigorous commit-reveal and benchmarking process, with rewards paid in SN14 tokens based on the measured performance improvement. The project operates as a Bittensor subnet and plans to deploy winning configurations as production inference endpoints, ultimately building an intelligent routing layer for optimal model serving.