2 ms·
Making models smarter and cheaper at the same time
https://arxiv.org/abs/2607.14431
- Wetime 4mo agoByte-exact KV grafting stores verified reasoning on disk. Replaying it lets a frozen 12B LLM beat 31B models at 8,700x less energy.
loading thread…