And... done.
What did we learn about Qwen 3.5 122b tonight?
Well, Zig isn't its strongest language. Finally hit a blocker when it tried to reconcile file management functions, and it got confused between the API spec for std.io and std.Io. It ended up looping, unable to get it right.
The assignment this time, is a screenwriting app I've wanted to write. The idea is that it takes your presentational screenplay, and converts it into storyboards, and seedance prompting.
And the idea is that it can keep your creative AI film projects organized and breathing, while plotting out your camera work for you.
Also, I insisted that it was written in Zig, with an esoteric database format I invented, and the requirement was vague, where it had to decipher what I was telling it to do, and the scope was intentionally ambiguous with a lot of room for scope creep.
It managed planning well, better than both Opus 4.8 and Kimi K2.7. This plan is beautiful, comprehensive, well designed. Life affirming stuff.
But it hit the complexity wall with Zig 0.1.6 internals.
Verdict: Overall, the strongest performance I've ever seen a local model accomplish. It did hit the complexity wall, although not in the places I expected, and it did fail on the spec. But... that was sort of the plan all along. It still got 30% of the way through the app and put in a respectable performance.