“The strategy of porting reasoning capabilities into AI2's existing OLMO recipe resulted in a model with a lower performance ceiling compared to rebuilding the DeepSeek R1 recipe.”