Re:Qwen4-Max
by bitexploder on Monday November 16, 2026 @07:38AM
(#50516959)
Yes, if you pay close attention to what they infer. Just click [-] and move on.
[ Parent
| on HN ]
Re:Qwen4-Max
by windexh8er on Monday November 16, 2026 @08:45AM
(#50517546)
I'm curious. What's the actual value, then? I fully understand minimizing a thread, but also am free to voice my opinion the same as you.
So, I'm curious... What should I be "paying close attention" to?
Re:Qwen4-Max
by bitexploder on Monday November 16, 2026 @09:12AM
(#50517786)
Alright, I was rude. So.. you can’t really benchmaxx pelicans specifically. You have to train the whole model on SVG and that is just generally useful. Secondarily it shows how a model understands the objects themselves and how to express it abstractly. IDK, it is a lot more interesting than assumed.
Here are a bunch of Pelicans I made in my basement. Side by side quant comparisons and something for fun :)
https://stealthgoat.com/fnfl/pelicans3.html https://stealthgoat.com/fnfl/pelicans8.html My favorite, and shows this is not all benchmaxxing, this model is just good at SVG: https://stealthgoat.com/fnfl/pelicans6.html Oh, and, quantization matters: https://stealthgoat.com/fnfl/pelicans2.html
Heres some examples.
Re:Qwen4-Max
by windexh8er on Monday November 16, 2026 @11:31AM
(#50519474)
Thanks for the explanation. I guess I still don't understand how this translates to other areas, outside of SVG generation, Pelican images and then bike geometry and build.
I'm curious if an increase in ability to generate this parlays into the model, generally, being better overall or in other adjacent areas?
Thanks again.