Seems promising, but it has a bit of the Apple M-series hype to it, in that yes, it can run large models, but the (shared) memory bandwidth isn't what you'd like it to be to do so.

It will be interesting to see real-world inference benchmarks.