Very singularity. This is a standalone model that can be used to incentivize search tool use without calling out to a search engine. They claim a 14b parameter version gets better training results than actually using google search for the RL phase.

Stuff like this is super important: it democratizes skills acquisition for everyone in the future, and speeds up research. Pretty cool!