News Summary

• JANG (Jang Adaptive N-bit Grading) is an open-source quantization format and toolkit that makes large language models run on Apple Silicon at 2-bit precision while staying coherent.
• JANG classifies tensors by sensitivity and gives critical layers (attention) more bits while aggressively compressing the bulk (MLP/experts) The result: a 122B model fits in 46 GB of GPU memory and answers questions correctly.
• The GGUF equivalent for MLX — models stay quantized in GPU memory at full Metal speed.