AirLLM v4.0.0 Trains 125B Qwen3.8-Flash-Next Model Under 6GB of VRAM
Summary
AirLLM's v4.0.0 release adds training support for Qwen3.8-Flash-Next, a 125B model, under 6GB of VRAM, streaming frozen weights one decoder layer at a time while adapters stay on the GPU. The project, maintained by Gavin Li, says it also runs Kimi K3 (2.8T) in 3.72GB. The Apache-2.0 library installs via pip.
Key Points
- AirLLM measures Qwen3.8-Flash-Next at 5.95GB of VRAM end to end on a single RTX 4090, the project reports.
- Kimi K3 support requires transformers 4.56.x, since the project says its remote code does not load on 5.x.
- Qwen3.8-27B trains in roughly 2GB of VRAM at sequence length 512, per the project's training notes.