3 ms·

I think if they made this for Qwen3.8-Next it could fit in a single 5090?

• kennywinker 15d ago
  180b * 1.76 bits per weight = 39.6 gigabytes. Best you could realistically run in 32gb is like 28gb, or a 127B param model
  • jokethrowaway 14d ago
    Qwen3.8-Next, thanks to its new architecture, is quite fast even if part of it is streaming from disk
    • kennywinker 14d ago
      Totally. Any MoE model can have experts swapped in and out from disk or system ram. I only framed it this way because the question was about the model fitting in vram.