MACaQ: Open-source quantizer with custom vLLM kernels and lower low-bit KLD

·2026.11.13 21:59

Key point

The framework supports weight and activation quantization for LLMs, text-to-image, and text-to-video models while maintaining inference speed.

MACaQ is a fast, open-source quantization framework that reduces low-bit KL-divergence while maintaining model size and inference speed.
It provides drop-in custom kernels for vLLM and llama.cpp to enable fast low-bit serving and activation quantization.
The tool supports both weight and activation quantization across LLMs, text-to-image, and text-to-video models.
Benchmarks indicate MACaQ outperforms popular standard quants, such as Unsloth and Bartowski GGUFs, on low-bit KLD.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.