[fzysingularity]

You can always distill VLMs into much smaller / faster models that’s specific to your domain or use-case.

What’s the use-case and what kind of latency do you require?