Quantized reasoning models close the gap on math benchmarks

New 4-bit quantization recipes preserve chain-of-thought quality within 2 points of full precision on MATH-500.

Summarized by the Neuricon wire. Every claim links back to its primary source where available.