Accuracy Recovery Adapters
Hey, it would be cool if Minimax could create some Accuracy Recovery Adapters for the quantized version of the Minimax M2 model. I think some people want to run the model on very limited systems and use very high quantization, like i1-Q4_K_S. With this, the model can run with 98GB. The issue is that the performance degrades with this high quantization. However, I have seen multiple projects where people make Accuracy Recovery Adapters; I have seen it done for LLMs and for image models like Qwen image and Flux models. The idea may originate from an Apple paper. I also found this Reddit post about it. It's very interesting, and I think something like this could help the community. https://www.reddit.com/r/LocalLLaMA/comments/1mytbfz/accuracy_recovery_adapter_with_selfgenerated_data/ https://huggingface.co/ostris/accuracy_recovery_adapters/tree/main