Microsoft's ThinkBit Keeps Qwen3 Reasoning at 2 Bits, Beats EfficientQAT by 43 MATH-500 Points at 1.7B

Pashmina Cameron (Microsoft Applied Sciences) published ThinkBit on Oct 27, a strictly scalar 2-bit quantization pipeline that explicitly protects reasoning. Standard perplexity hides degradation — removing codeword search or recalibration keeps 7-task means within 1.5 points while MATH-500 crashes from 43.0 to 28-37. ThinkBit's 0.52GB Qwen3-1.7B scores 73.0 on sampled MATH-500 versus 30.0 for EfficientQAT and 42 for the best published 2-bit GGUF; the 2.08GB 8B version holds 92.2. Framework: Kronecker-Hessian bit allocation across 1.71/2.125/3.125-bit tiers, Gumbel-softmax codeword search, KL recalibration, ReinMax QAT, held-out MATH checkpoint selection.