Intel squeezed a 1.58-bit LLM down to 1.485 bits without changing a single weight

Intel researchers presented the BITCOS format, which stores ternary model weights more compactly than the theoretical minimum of 1.58 bits. One checkpoint was compressed to 1.485 bits per weight, speeding up decoding by up to 18% on CPU and 27% on GPU without retraining or loss of accuracy.

• BITCOS stores zero and non-zero weights separately: 2 − z bits per weight
• Compression is more efficient than standard packing when zero share exceeds 37.5%
• Decoding speedup: 10–18% on Xeon and 2–27% on Arc GPU
• Most sparse checkpoint Qwen3-1.7B: 51.48% zeros, 1.485 bits