Running a large language model is often bottlenecked less by compute than by memory: fitting the weights, moving them around, storing checkpoints. Squeezing each weight toward two bits helps, but the usual tricks stall. Simple rounding loses too much at that budget, while richer vector methods need explicit codebooks and index lookups that add overhead.

BiSCo-LLM takes a codebook-free route. It maps small weight chunks onto a unit sphere and turns them into compact binary sign codes, so most of what's stored is a packed stream of bits. A second stage encodes the leftover error, and a short distillation step re-tunes the model to cover the mismatch, with a few sensitive channels kept at higher precision. Usefully, the reported storage budget counts everything: codes, decoders, adapters, and metadata.

Despite the neural decoding label, this decompresses model weights, not brains. From the abstract, so check the paper for real compression ratios and accuracy.