‹ BackHN Continuity

Thread

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

741 points · 330 comments · snehesht

  1. soltanov · · focus · HN ↗
    Does this performance survive a fixed task suite and harness when accuracy, energy, long-context behavior, and run-to-run variance are measured alongside tokens per second?
Open on Hacker News to reply ↗

Unofficial Hacker News client; not affiliated with Y Combinator.