New open-source method fine-tunes big MoE models on consumer GPUs
A new method called USAF has been released. The idea is simple: if a GPU can run inference on a model, it should also be able to fine-tune it (retrain it further for a specific purpose). This method makes it possible to fine-tune Qwen3-30B-A3B, a large MoE (, a model made of several smaller expert sub-models) model, on an ordinary consumer GPU like an AMD RX 6750 XT with just 12GB of memory.
Instead of using adapters (small extra modules commonly attached for training), it works by sparsely training only some of the expert weights and the router (the part that decides which experts to use) directly. The project is fully under the , and its creator says it isn't meant to be a business or monetized product.
Key points
- USAF is a new sparse training method for MoE models on consumer GPUs
- Demonstrated Qwen3-30B-A3B on an AMD RX 6750 XT with only 12GB of memory
- Trains select expert weights and the router directly instead of using adapters
- Released fully on GitHub under the
- Creator states the project isn't intended for commercial or purposes