I uploaded Qwen 3.8 27B with SGLang + DFlash 2 on 4 H100s, and I was excited to see 6K decodes per second with 48 parallel connections.
I also tried Jamidusu's Myeongri consultation using Horncheon's CLI, and it was good because there were no noticeable errors and it was fast.
GLM 5.3 absolutely cooked Qwen3.8 Max 😭
Qwen 3.8 Max returned basically a black screen And somehow Qwen cost me ~11× more for the generation GLM 5.3 is looking seriously impressive right now
Qwen 3.8 27B vs Ornith-1.5-35B 🔥
Top: low-poly pirate ships Bottom: NES circuit racers Qwen takes both. Pirate → smoother orbits + cleaner details NES → full mini-game (title → proper track → 1st place results) Ornith’s racer feels like a prototype (frozen timer +
Ornith-1.5-35B-A3B-oQ8e-mtp in action with oMLX v0.6.3rc1 Video normal speed!
🚀 Model used here: Tomorrow I'll compare results vs Qwen 3.8 27B!
You have reached the end of the archive
All of qwen38