"The 31 billion model scored 85.7% on the GPQA diamond benchmark, which is a pretty tough scientific reasoning test, and it ranks third among open models under 40 billion parameters."