Thread

Clef: Open-weight decision models, and new RL fine-tuning platform

637 points · 217 comments · jasondavies

bityard · 1790874486 · focus · HN ↗
Clef is based on Qwen3.8-27B and Clef-flash is based on Qwen3.8-9B (edit: actually Qwen3.5-9B). So, similar in spirit to Kev by my understanding, but based on a newer model.
1. NitpickLawyer · 1790880016 · focus · HN ↗
  > and Clef-flash is based on Qwen3.8-9B
  There is no official qwen 3.8 9b
  From the model card:
  > Clef-Flash is post-trained from Qwen/Qwen3.5-9B. See Clef for the larger variant.
  1. ddarolfi · 1790880498 · focus · HN ↗
    It's based on Qwen3.5-9B, maybe a typo
    verdverm · 1790895556 · focus · HN ↗
    I've been guilty of the same wishful projection
  2. bityard · 1790881451 · focus · HN ↗
    Thanks, I missed that. Fixed my comment.
2. okpatil · 1790881581 · focus · HN ↗
  Atom is 60M Param (around 133x to 400x smaller).
  16ms latency. And locally run.
  <a href="https://at0m.pienomial.com/" rel="nofollow">https://at0m.pienomial.com/
  Why go big when you can go small ?
  1. kamranjon · 1790883967 · focus · HN ↗
    Cause it's not open?
    okpatil · 1790884261 · focus · HN ↗
    Good point.
    To counter, most of the AI is not open. So is none of Microsoft Products. As long as they work, we keep using them.
    verdverm · 1790895659 · focus · HN ↗
    counter point, I've stopped using all closed models and harnesses as a life choice
    Ai is too important and transformational to let Big Ai dominate in a closed ecosystem, thankfully the Chinese have a different mindset and approach
    okpatil · 1790957451 · focus · HN ↗
    That's an absolutely fair way. More power to you !
  2. ricardobeat · 1790932435 · focus · HN ↗
    Because with a tiny model you're skipping all the intelligence and world knowledge that makes it useful without fine tuning. `typed-decisions` is almost entirely text classification tasks.
    okpatil · 1790957344 · focus · HN ↗
    Don't assume. It has generic world knowledge. It is performing well on the benchmarks we didn't even train it on.

Open on Hacker News to reply ↗

Unofficial Hacker News client; not affiliated with Y Combinator.