Neural.labs releases QuantForge: open-source 3-bit quantization tool for deploying design models

Tech · 5 min read

QuantForge includes an end-to-end pipeline: calibration routines for generative and discriminative heads, automated per-layer quantization strategies, and a fallback scheme that keeps sensitive layers at higher precision. The tool is compatible with common runtimes and supports export to coreml, ONNX, and custom accelerators.

The company demonstrated QuantForge on several design-focused models (layout detectors and small image encoders), showing significant reductions in memory footprint and improved CPU performance while preserving layout fidelity and reducing hallucinations. The project includes evaluation scripts for downstream metrics like bounding-box stability and token alignment.

Neural.labs released QuantForge under an MIT-like license for community adoption and encourages maintainers of design tools to integrate the tool to enable local and offline model usage. The effort responds to increasing demand for privacy-preserving, fast inference in creative tooling.