The small-model numbers are pretty serious too: IFM says the 7B can now handle AIME competition math, a class of work that required 100B+ models roughly 1 year ago, while the 0.9B, 3.7B and 7B results are also reflected in independent Artificial Analysis evaluation.

The 32B ranks among the top dense models below 40B, while the 375B-A23B ranks among the top models below 400B across general, reasoning, coding and agentic evaluations.

IFM also says the 375B-A23B costs roughly as much to run as a 25B dense model, which is where the sparse design starts getting very interesting.