Tag
autotrust 发布的 GLM5.3-Flash-E224-DGX-Spark 模型仓库，基于 zai-org/GLM-5.3-Flash 的图像-文本到文本（image-text-to-text）量化版本，支持 NVFP4/8-bit 量化，可在 DGX Spark (GB10/Blackwell) 上通过 vLLM 运行。
ISTA-DASLab released a capability-targeted compressed version of Qwen3.8-Flash-Next that prunes half the routed experts and stores remaining weights at 3.5 bpw, shrinking the 354 GB model to 29.6 GB so a 176.9B-parameter model fits in a single 32 GB accelerator, while retaining 98.7% of LiveCodeBench v6 and 91.3% of SWE-bench Verified scores.
This paper describes six submissions to the WMT26 Model Compression Shared Task, using routing-informed expert pruning and MXFP4 quantization to compress GPT-OSS-20B into smaller translation models with parameters ranging from 4.186B to 7.770B.
This paper proposes using router weight sensitivity under lightweight fine-tuning (e.g., LoRA) to identify and prune experts in Mixture-of-Experts models, enabling significant memory and latency reductions with minimal accuracy loss.
This paper introduces LorExperts and BTExperts, router-preserving compression methods for Mixture-of-Experts LLMs that cluster experts and represent non-dominant members as low-rank corrections, improving compression quality over prior methods like D2-MoE.
jabbatheduck released a GGUF quantized version of the REAP expert-pruned DeepSeek-V4-Flash checkpoint, aggressively compressed for memory-constrained inference on consumer GPUs while preserving router and attention precision.
This is a highly experimental GGUF version of the 2.8T-parameter Kimi K3 MoE model, with 55% of experts pruned and quantized to ~2.15 bpw (319 GiB). It requires a specific llama.cpp PR and custom patches to run, and includes detailed instructions for usage.
SHAPE proposes a coalition-aware expert pruning framework for sparse MoE LLMs that uses Shapley-style attribution over routing traces to identify essential experts, achieving competitive accuracy under 20-40% pruning and reducing GPU memory footprint.
ConMoE proposes a train-free prototype remapping framework for Mixture-of-Experts (MoE) compression, which selects a subset of experts as reusable prototypes and deterministically remaps original expert calls to them, reducing memory usage without weight updates or fine-tuning.