Respectfully, I'd have to disagree, the models are extremely filtered, it seems like they have almost no real world knowledge and are likely trained directly on o3/o4 mini outputs (synthetic data, like the Phi series). Even then, they're quite bad at code, at least the basic website frontend stuff (which they're supposed to be good at) and other things.
I think they might only be good on very specific math and (scientific?) programming tasks only in popular languages. Some people have speculated that there is actually no "base" pretrained model and the whole model was trained from scratch on outputs from other OpenAI models.
It should still be a decent model for doing tool calls and being a basic "agent" of course, but so far it doesn't seem to be a breakthrough at all.
As an /lmg/ poster puts it succinctly:
>just saw someone elsewhere say that the model is just Phi 5, and I think that's the best way of putting it
>feels brittle in exactly the same way as the Phi series, so benchmaxxed and synthetic that it disintegrates when given anything even slightly OOD
Related ongoing thread:
Open models by OpenAI - https://news.ycombinator.com/item?id=44800746 - Aug 2025 (475 comments)
specifically here: https://news.ycombinator.com/item?id=44804034
Respectfully, I'd have to disagree, the models are extremely filtered, it seems like they have almost no real world knowledge and are likely trained directly on o3/o4 mini outputs (synthetic data, like the Phi series). Even then, they're quite bad at code, at least the basic website frontend stuff (which they're supposed to be good at) and other things.
I think they might only be good on very specific math and (scientific?) programming tasks only in popular languages. Some people have speculated that there is actually no "base" pretrained model and the whole model was trained from scratch on outputs from other OpenAI models.
It should still be a decent model for doing tool calls and being a basic "agent" of course, but so far it doesn't seem to be a breakthrough at all.
As an /lmg/ poster puts it succinctly:
>just saw someone elsewhere say that the model is just Phi 5, and I think that's the best way of putting it
>feels brittle in exactly the same way as the Phi series, so benchmaxxed and synthetic that it disintegrates when given anything even slightly OOD
>the ultimate small model smell
This is exciting (re openai/gpt-oss-20b):
> First impressions: this is a really good model, and it somehow runs using just 11.72GB of my system RAM.
So could maybe run this in my base M4 mini?
So do these best the offerings from Qwen, Moonshot and Z.ai as simonw implies/predicts? And with only 16G needed!!?
[dead]