Conversation

Notices

1. Embed this notice
   Blaise Pabón - controlpl4n3 (blaise@hachyderm.io)'s status on Wednesday, 07-May-2025 01:48:30 JST Blaise Pabón - controlpl4n3

   in reply to

   • Zach Bennoui
   • Chi Kim

   @chikim @ZBennoui
   I have 256G, but only 20 cores and no GPU, does MoE still help in that case?

   • Embed this notice
     Chi Kim (chikim@mastodon.social)'s status on Wednesday, 07-May-2025 01:48:31 JST Chi Kim

     in reply to

     • Zach Bennoui

     @ZBennoui Yea try them on local machine. Especially if you have enough ram, qwen3:30b is really fast because of MoE architecture.

   • Embed this notice
     Zach Bennoui (zbennoui@dragonscave.space)'s status on Wednesday, 07-May-2025 01:48:32 JST Zach Bennoui

     in reply to

     • Chi Kim

     @chikim I tried the really big model on the HF demo, too scared to make an account on the QWEN Chat site lol. Haven't yet tried the smaller local versions yet but the big one is really quite good.

   • Embed this notice
     Chi Kim (chikim@mastodon.social)'s status on Wednesday, 07-May-2025 01:48:33 JST Chi Kim

     Qwen3 is released right before LlamaCon tomorrow! lol 32K context length, tool calling, a way to turn on/off reasoning with /think /no_think in prompt, 119 languages support. 6 dense models (0.6b, 1.7b, 4b, 8b, 14b, 32b) and 2 MoE models (30b-a3b, 235ba22b). #LLM #AI #ML https://qwenlm.github.io/blog/qwen3/