Jangq

JANG is an open source quantization method for running larger AI models on Apple Silicon Macs. It is built around the MLX framework and shrinks model files so that bigger, more capable models fit in the memory of a laptop or desktop Mac while keeping output quality higher than uniform quantization at the same size. Standard quantization gives every tensor the same bit width. JANG instead sorts tensors into sensitivity tiers and assigns bit widths accordingly. Attention layers, which the vendor says make up about 12 percent of parameters, receive 5 to 8 bits, while MLP layers are compressed to 2 to 4 bits, at an average overhead of about 0.3 extra bits. The vendor publishes a curated benchmark page comparing JANG builds with MLX baselines. Its headline examples are MiniMax M2.5 at 82.5 GB scoring 74 percent on a 200 question MMLU sample against 26.5 percent for MLX 4 bit at 119.8 GB, and Qwen3.5 122B at 44.7 GB scoring 79 percent against 46 percent for an MLX mixed 2 and 6 bit build at 45 GB. The vendor states that it lists only comparisons where JANG wins and uses fewer bits or less memory. It suits developers and enthusiasts who run local language models on Macs and are limited by unified memory. The project is released under the Apache 2.0 license, with source on GitHub. The site describes it as free, and no paid plans are published.