AQLM

AQLM is an open-source PyTorch library published on GitHub for extreme compression of large language models via additive quantization, together with its successor method PV-Tuning. The project has gathered around 1,300 GitHub stars and is actively maintained, with updates in 2026.

Overview

Quantizing an LLM to 2 bits per weight normally destroys accuracy. AQLM — Additive Quantization for Language Models — instead learns multiple codebooks per weight group, representing each block of weights as a sum of codebook entries, which preserves far more quality at extreme compression ratios. The repository is the official implementation of the ICML 2024 paper and of PV-Tuning, a follow-up that fine-tunes the quantised model beyond straight-through estimation for even better results. It ships quantised checkpoints of popular models, calibration and fine-tuning code, and fast inference kernels, letting a model that normally needs a datacenter GPU run in a fraction of the memory. For practitioners, it is one of the reference codebases for pushing LLMs onto smaller hardware without retraining from scratch.

Key Features

• Extreme LLM compression down to ~2 bits per weight
• Additive multi-codebook quantization learned per weight group
• PV-Tuning fine-tuning for additional accuracy recovery
• Ready-made quantised checkpoints of popular open models
• Calibration and evaluation pipelines included
• Fast inference kernels for the quantised format

💳 Pricing

AQLM is free and open source under the Apache 2.0 license. The only cost is the compute you quantise and run on.

Who Is It For

ML engineers and researchers who need large models to fit on limited GPU memory, and anyone studying state-of-the-art weight quantization.