You lost me at calling proponents of the post-compression movement "haters".
Vincent Arsenault
AI & ML interests
Recent Activity
Organizations
Export script
Best way to finetune for structured final output instead of plain text final output ?
Format conversion onnx, rknn
Pytorch and onnx format.
Updated procedure for ONNX export
God help the people you advise on AI matters.
Dude, sure, make America unilingual again.
Still, of course, there are absolutely no indication at all that multilingual training yields better results even at unilingual tasks. None at all.
Also, you do know that Hugging Face is a French company, right?
Barely, but I certainly didn't know the businesses there were required by law to only serve white-speaking customers.
[...] if unilingual (therefore small) model is capable of doing a niche task [...]
Ironically, niche is a french word that means the exception, i.e: not the real world.
And I don't think Smol models are meant to be niche, no.
it's even reasonable to translate the original
This is a VLM so again, no. You don't translate images.
Personally, I don't think it's fair to compare a unilingual model on unilingual benchmarks against multilingual models. Are unilingual models that relevant in 2025, anyways? There are no use cases in the real world for unilingual models.
MMBench Leaderboard
Explore MMBench Leaderboard data
SEED-Bench Leaderboard
Submit model evaluation results to leaderboard
Open VLM Leaderboard
VLMEvalKit Evaluation Results Collection
Vision Arena (Testing VLMs side-by-side)
Explore Vision Arena’s computer‑vision tools online
Vidore Leaderboard
Browse and compare visual document retrieval model scores