AnalysisAI ModelsSeptember 20, 2026

Qwen3.8-Flash-Next builds browser game in one shot on 4x V620

A Reddit user ran Qwen3.8-Flash-Next locally in Intel AutoRound W4A16 on 4x V620 GPUs at roughly 2k prefill and 70 tokens/sec decode for about 3 hours. The model ran two browsers simultaneously, testing and fixing its own Cosmic Arcade game.

1 source

More stories today

Open the live feed