The authors built ThinkBit for Qwen3-1.7B and Qwen3-14B.
Scalar two-bit weight quantization can preserve language model reasoning accuracy
ThinkBit, a five-stage framework, keeps MATH-500 accuracy within a few points of bfloat16 on models up to 14B while shrinking memory traffic by up to 8x.
Big Tech
Rhea Sanjay Sukthanker · Pashmina Cameron
Microsoft Applied Sciences Group
Research Digest··3 min read
The authors present ThinkBit, a five-stage framework for quantizing reasoning language models to two bits per weight in a strictly scalar format, where each weight is an integer codeword times a bfloat16 scale shared by 128 weights.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.