Show HN: Ported Cerebras Reap to MLX – Prune Moe Experts on a MacBook (github.com)

A new tool called REAP MLX has been released, enabling expert pruning for MLX-LM mixture-of-experts (MoE) models specifically on Apple Silicon Macs. This software applies innovative Router-weighted Expert Activation Pruning (REAP) to improve the efficiency of MLX-LM models without needing complex setups like CUDA or PyTorch. REAP MLX streamlines the pruning process by calibrating the model, observing activation metrics, and selectively removing low-saliency experts—all while ensuring that the output model's integrity is maintained through detailed validation telemetry. This development is significant for the AI/ML community as it democratizes access to advanced pruning techniques, allowing researchers and developers to conduct local experiments more easily. With a focus on accessibility, the software supports a direct command-line interface and operates in a lightweight import-light package, making it user-friendly for those without extensive computational resources. The output includes structured telemetry data that can guide further research and optimization strategies in model deployment, ultimately enhancing the performance and scalability of AI solutions in real-world applications.