MLX-compatible REAP for pruning MoE models on Apple Silicon
-
Updated
Nov 4, 2026 - Python
MLX-compatible REAP for pruning MoE models on Apple Silicon
MoE expert compression toolkit — prune, merge, and forge specialist models for local inference.
RAZOR: Pruning Replaceable Experts in LLMs — training-free expert pruning for Mixture-of-Experts (MoE) models, with the MoECal calibration corpus
Experimental MoE expert pruning and router analysis for safetensors checkpoints. CPU probing, keep plans, dry runs, and checkpoint rewriting with synthetic tests.
Less is MoE (EMNLP 2026 Main); Fisher-MoE: intra-expert structured pruning and compression for Mixture-of-Experts (MoE) LLMs. Prunes FFN intermediate dimensions by Fisher importance instead of dropping whole experts. Official code.
DeepSeek-V4.1-Flash-1017 pruned to 192 of 384 experts per layer, at ~2.85 bpw with Engram read from NVMe, served on one Jetson AGX Thor. Weights on Hugging Face; full pipeline and gates here.
PLI Labs: legal model prunes (REAPs), owned-hardware serving research, evaluation methods, and open-source tools from Proprietary Legal Intelligence.
Expert pruning of two MoE models for one owner's traffic, so both fit on one DGX Spark: DeepSeek V4 Flash 95 to 70 GiB with no loss in review, and Qwen3.6-35B-A3B at 208 of 256 experts with vision, side by side. A finished record, with the negative results.
Code release for DIET: DeletIon-response Expert Trimming for Video Diffusion Transformers
KL-divergence + blind task eval of GLM-4.7-Flash and its REAP-pruned variant vs gpt-oss-20b on llama.cpp
Run Sarvam's 105B FP8 MoE decision model (Prism) on a laptop iGPU by streaming layers from SSD, plus an expert-pruning study. Pure PyTorch.
GLM-5.3-Flash (321B MoE) pruned to 144 of 288 experts and quantized to NVFP4/GGUF for one Jetson AGX Thor. Layer-streamed pipeline, paired teacher evaluation, and the negative results.
REAP pruning + NVFP4 quantization of Ornith-1.5-35B-A3B for 16GB VRAM
HOPE expert pruning of MiMo-V2.6-Flash (309B omnimodal MoE) to 128 of 256 experts for one Jetson AGX Thor, calibrated with real image, audio and video. Published; not yet evaluated.
How I fit 35B-parameter MoE models into 16 GB of consumer VRAM: REAP expert pruning + NVFP4 quantization on an RTX 5070 Ti (SM120), served on vLLM, with measured evals.
Serving a large MoE coder model on consumer Blackwell GPUs via REAP expert pruning + NVFP4 quantization.
Shrinking Qwen3.5-35B-A3B on one 16 GB GPU: MoE expert pruning (REAP), router distillation and GGUF quantization, with interactive lessons and every command
To associate your repository with the expert-pruning topic, visit your repo's landing page and select "manage topics."