MLX-compatible REAP for pruning MoE models on Apple Silicon
-
Updated
Nov 4, 2026 - Python
#
expert-pruning
Here are 17 public repositories matching this topic...
MoE expert compression toolkit — prune, merge, and forge specialist models for local inference.
machine-learning pytorch moe pruning quantization model-compression mixture-of-experts llm llama-cpp vllm local-llm llm-inference gguf expert-pruning
-
Updated
Nov 3, 2026 - Python
RAZOR: Pruning Replaceable Experts in LLMs — training-free expert pruning for Mixture-of-Experts (MoE) models, with the MoECal calibration corpus
pytorch moe pruning model-compression mixture-of-experts llm training-free llm-compression expert-pruning moe-pruning
-
Updated
Oct 1, 2026 - Python
Experimental MoE expert pruning and router analysis for safetensors checkpoints. CPU probing, keep plans, dry runs, and checkpoint rewriting with synthetic tests.
python pytorch moe model-compression mixture-of-experts llm safetensors qwen mixtral deepseek router-analysis olmoe expert-pruning
-
Updated
Oct 11, 2026 - Python
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.
moe pruning model-compression efficient-inference fisher-information mixture-of-experts neuron-pruning structured-pruning large-language-models llm vllm qwen llm-compression emnlp2026 olmoe expert-pruning
-
Updated
Sep 25, 2026 - Python
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.
-
Updated
Oct 31, 2026 - Cuda
Code release for DIET: DeletIon-response Expert Trimming for Video Diffusion Transformers
-
Updated
Sep 30, 2026 - Python
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.
-
Updated
Oct 26, 2026 - Python
KL-divergence + blind task eval of GLM-4.7-Flash and its REAP-pruned variant vs gpt-oss-20b on llama.cpp
moe glm quantization kl-divergence runpod llama-cpp local-llm gguf llm-evaluation gpt-oss expert-pruning
-
Updated
Jun 12, 2026 - Python
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.
moe quantization reap mixture-of-experts blackwell vllm llm-inference consumer-gpu nvfp4 expert-pruning
-
Updated
Nov 8, 2026
PLI Labs: legal model prunes (REAPs), owned-hardware serving research, evaluation methods, and open-source tools from Proprietary Legal Intelligence.
open-source evaluation inference reap legal-ai expert-pruning pli-labs proprietarylegal owned-hardware
-
Updated
Nov 3, 2026 - Python
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.
-
Updated
Oct 26, 2026 - Python
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.
pytorch offloading model-compression technical-report mixture-of-experts on-device-ai xpu fp8 intel-arc llm-inference sarvam expert-pruning
-
Updated
Nov 8, 2026 - Python
REAP pruning + NVFP4 quantization of Ornith-1.5-35B-A3B for 16GB VRAM
-
Updated
Sep 22, 2026 - Python
Serving a large MoE coder model on consumer Blackwell GPUs via REAP expert pruning + NVFP4 quantization.
-
Updated
Sep 22, 2026 - Python
Expert pruning of DeepSeek V4 Flash for one agent's traffic on one DGX Spark: 160 of 216 routed experts per layer, 5 experts per token, 74 GiB in place of 95 GiB, with no loss in the final review.
moe model-compression mixture-of-experts llm vllm llm-inference speculative-decoding deepseek dgx-spark deepseek-v4 expert-pruning exl3
-
Updated
Nov 12, 2026 - Python
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
benchmarking machine-learning pytorch moe pruning quantization knowledge-distillation interactive-tutorial model-compression reap mixture-of-experts llm llama-cpp qwen gguf imatrix qwen3-5 expert-pruning
-
Updated
Oct 24, 2026 - HTML
Add this topic to your repo
To associate your repository with the expert-pruning topic, visit your repo's landing page and select "manage topics."