Tencent/AngelSlim

Supports multiple compression strategies—quantization algorithms (FP8, INT4, INT8, exotic formats like NVFP4 and 1.25-bit Sherry), speculative decoding frameworks (Eagle3, SpecExit), and pruning—across LLMs, vision-language models, and diffusion models. Built on a unified post-training quantization (PTQ) pipeline optimized for single-GPU operation on models up to 235B parameters. Integrates with Hugging Face and ModelScope ecosystems, with inference backends including vLLM and Torch.

536 stars and 5,117 monthly downloads. Actively maintained with 21 commits in the last 30 days. Available on PyPI.