The A.I. Daily
Lean-QwenFlash Runs 125B Model on Low Memory
A developer adapted the glm-flash-lite SSD offloading approach to Qwen3.8-Flash-Next (125B), creating Lean-QwenFlash, which decodes at ~20 tok/s with 16GB VRAM and 8GB RAM. An RTX 4090 hits 20 tok/s decode and 1,627 tok/s prefill at 262k context; an RTX 3080 laptop achieves 18-21 tok/s at 100k context. Code and scripts are open-sourced on GitHub.