What's your favorite medium to large (~50-150b params) MOE model? Aside from Qwen flash next.
I recently got some more RAM on my desktop, which has opened up running larger MOEs by splitting them into RAM/VRAM.
Proof1 community thread
Key points
- I know everyone loves Qwen flash next, but I'm curious what other ones are worth trying.
- Qwen 3.5 122b A10b (Is this even worth trying or is it totally supplanted by flash next?)
- I know flash next is the king there, and realistically I just use Claude or Codex when I want coding.
- So if a model sucks at coding but is great at something else, I'm interested!
Sources (1)
- [1]What's your favorite medium to large (~50-150b params) MOE model? Aside from Qwen flash next.r/LocalLLaMA (top, daily) · Oct 22, 03:44 AM
I recently got some more RAM on my desktop, which has opened up running larger MOEs by splitting them into RAM/VRAM.
I know everyone loves Qwen flash next, but I'm curious what other ones are worth trying.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 22, 2026I have an answer! Qwen Flash Next IQ2 IS better than Q8 of 27B (for me).
- Oct 21, 2026Qwen researching Qwen
- Oct 21, 2026OpenAI Frontier Models Arrive on Google Cloud and Model Garden
- Oct 20, 2026Qwen3.8-Flash-Next vs Opus 5.5: Six-way benchmark on a real Rust feature with Claude Code and Pi
- Oct 20, 2026How I built my ‘everything’ app
- Oct 20, 2026Voxtral 3
