Intel Squeezed a 1.58-Bit LLM Down to 1.485 Bits Without Changing a Single Weight
The New Stack, Thursday, September 17th, 2026
The New Stack reports Intel's BITCOS format compresses ternary LLM weights below 1.58 bits and speeds GPU decoding by up to 27%.
Intel researchers developed BITCOS, an encoding format for ternary large language model weights that gets below the theoretical 1.58-bit floor without altering any weight values.
The trick is exploiting the zero-heavy distribution typical of ternary quantized models, packing the resulting skew more efficiently than a naive encoding. The New Stack reports the format reaches 1.485 bits per weight while also improving decoding throughput on GPUs by as much as 27%.
Because it is lossless with respect to the original ternary weights, it is a drop-in win for memory-bound inference rather than another accuracy trade-off.