Re:DeepSeek V4 Flash 0731
by kamranjon on Friday August 07, 2026 @05:22PM
(#49216390)
This is pretty interesting, I've never heard of this approach before - do you know if there is a research paper that covers how this was achieved?
[ Parent
| on HN ]
It’s an adaptation of CoLaR, but my implementation is a little different:
- Dedicated stop head to fire when latent thinking hits threshold
- MTP support with training taking draft support as first class
- different architectural layer 35 -> layer 42 writeback. So latents skip roughly 6.2 tokens of reasoning per token, then never make it to decoded output