7 ms·
Perhaps it might be interesting: a latent thinking version is here https://huggingface.co/nmitchko/DeepSeek-V4-Flash-0731-Latent-Reasoning https://huggingface.c
by nmitchko 1mo ago
Perhaps it might be interesting: a latent thinking version is here https://huggingface.co/nmitchko/DeepSeek-V4-Flash-0731-Latent-Reasoning https://huggingface.co/nmitchko/DeepSeek-V4-Flash-0731-Laten...
Does no thinking emissions for context saving.
- krackers 1mo agoThis deserves its own hn post!
- nmitchko 1mo agohttps://news.ycombinator.com/item?id=49230550 https://news.ycombinator.com/item?id=49230550
- kamranjon 1mo agoThis 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?
- nmitchko 1mo agoIt’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