Samuel
oh I mean I’m not the blindsided kind of surprised, I knew it was coming, but I’m still suprised that it actually happened
1 4
philpax
word on the street that they're on the verge of dropping GPT-5, so it doesn't matter if they cannibalise their own offerings today
3
symmetricalboy
The open model ecosystem lives alongside the closed models perfectly fine. they don't conflict or detract. neither is wrong or better, imo, & they serve very different purposes. they have their SaaS options to bank on & they also ship some OSS. nothing new happening here at all. same ol' internets.
2
Yin Shen
It's not a SOTA model. Also, OpenAI has been losing researchers to Meta. Seems to be a play at both market dominance (denying competitors revenue) and reputation repair (staff retention)
Eris ⌬
They do, unfortunately, seem to perform horribly. Hallucinations are INSANE, and the 120B one does worse on SVGBench than o4-mini (which itself is below most of the newer Qwen models)
1 6
Eris ⌬
If it released back at the start of June, it would've been SOTA for Open models... but unfortunately things have moved on from there and the performance + hallucinations make these pretty useless, even for the cheap prices the APIs cost. Maybe they turn out to be very finetunable but I'm not betting
1 3
Monty
“As part of today’s release, Microsoft is also bringing GPU-optimized versions of the gpt-oss-20b model to Windows devices. Powered by ONNX Runtime, these models support local inference” Ah, there we go. Tough to keep your model private on the edge anyway, so this makes them look better.
1
mussar
imo its worth playing with but possibly a "too little, too late" situation unless it is an order of magnitude better than existing OSS models.
Bumping into people
Ahh.. Why release a lower performing model for OSS? It just feels irresponsible to waste energy training a sub-par model when there are already ones better out there. Maybe I am missing something entirely?