‹ BackHN Continuity

Thread

Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s

741 points · 330 comments · snehesht

  1. zkmon · · focus · HN ↗
    >> The model is a team of 24,576 small specialists ("experts")

    That's a neat number (576 is the square of 24). Ofcourse it must have come from 24 * 2^10.

    1. kzrdude · · focus · HN ↗
      I have to say that "team of specialist" and "model of experts" as explanations and names give a quite misleading picture about how it works, at least I thought so when I learned about how it worked.
Open on Hacker News to reply ↗

Unofficial Hacker News client; not affiliated with Y Combinator.