Cacheon
탈중앙화 LLM 추론 최적화 경쟁 플랫폼
공연
RD 인기지표
RD 성장지표
비교
팔로우 목록
인물
Cacheon 소개
Cacheon is an inference-optimisation subnet on Bittensor, subnet 14, which turns faster serving of a fixed model into a measurable contest. Miners submit containerised inference servers; validators run them against a pinned vLLM baseline on identical hardware and load, measuring time to first token and decode throughput, then checking outputs against the reference model with log probabilities. Only fast and correct servers earn weight.
Scoring credits the smallest attributable change against the incumbent: each candidate runs as a complete isolated engine in separate processes, with hardware lanes swapped between baseline and candidate so a repackaged earlier win cannot pass as new, and copies of the reigning champion are disqualified. Rewards are winner-take-most. From May 2026 evaluation rounds began publishing decode speeds on standardised hardware.