>> [...] achieving state-of-the-art performance in mathematical and *coding tasks* [...]

And it fails at one of the simplest coding tasks where a Granite model at nearly half the size has no problems.

It's probably an important discovery but seemingly only usable in an academic context.