What happened to Llama 3.2 90b-vision?

My own experience with it was pretty bad, they attempted to bake in way too much censorship in it. It failed even basic tests some YouTubers thrown at it specifically due to degradation caused by overcensoring: [https://www.youtube.com/watch?v=lzDPQAjItOo](https://www.youtube.com/watch?v=lzDPQAjItOo) . For vision tasks, Qwen2-VL 72B is better in my experience. I can run it locally using [https://github.com/matatonic/openedai-vision](https://github.com/matatonic/openedai-vision) . It is not as VRAM efficient as TabbyAPI, so requires four 24GB GPUs to run it and even that feels like a tight fit. And it still not as good as text-only Qwen2.5 or Llama3.1, and loading vision model takes few minutes, then few more minutes to get a reply, and load back normal text model, so currently large vision models are not very practical. My guess, for heavy vision models to become more popular, they need to become more widely supported by popular backends such as Llama.cpp or ExllamaV2, but there are a lot of challenges to implement vision model support. This way, they would become more VRAM efficient and may gain better performance, and when we have good vision models that also remain good at text-only tasks, it may become more practical to use them. I still use vision models quite often, but I understand why they are currently not very popular due to issues mentioned above.