Making models smarter and cheaper at the same time
2 pointsby Wetime18 hours ago1 comment
  • Wetime18 hours ago
    Byte-exact KV grafting stores verified reasoning on disk. Replaying it lets a frozen 12B LLM beat 31B models at 8,700x less energy.