TechPrufer
Get started
Dropbox
Dropbox logoDropbox · Feature trail

LLM Quantization Optimization

2 posts · Oct 2025

What this trail covers

Dropbox's Mobius Labs team developed Half-Quadratic Quantization (HQQ), a novel data-free quantization technique for large machine learning models, particularly LLMs. HQQ significantly speeds up the quantization process (e.g., 50x faster for Llama-2-70B compared to GPTQ) while achieving competitive compression quality. It minimizes errors in model weights by incorporating a sparsity-promoting loss and using a Half-Quadratic solver with closed-form solutions, enabling inference in half-precision.

All Dropbox trails
  1. With Mobius Labs' Aana models, we're bringing deeper multimodal understanding to Dropbox DashThis post details the integration of Mobius Labs' Aana multimodal models into Dropbox Dash. It highlights Aana's architecture, which combines open-source foundation models for speech, vision, and language, and emphasizes its efficiency through advanced inference optimizations like HQQ (Half-Quadratic Quantization) for low-bit inference and Gemlite for accelerating core AI operations. The Aana SDK is also mentioned for orchestrating these optimizations and enabling flexible deployment of multimodal workflows. The post explains how these technologies enable deeper multimodal understanding of rich media content at scale within Dropbox Dash.
  2. Half-Quadratic Quantization of large machine learning modelsThis post introduces Half-Quadratic Quantization (HQQ), a novel data-free quantization technique for LLMs. HQQ focuses on minimizing errors in model weights by incorporating a sparsity-promoting loss (l_p-norm) and using a Half-Quadratic solver to optimize zero-point and scaling parameters. This approach achieves significant speed-ups (over 50x faster than GPTQ for Llama-2-70B) and competitive compression quality without requiring calibration data, enabling faster deployment of large models.