mradermacher

/

model_requests

like 107

I queued them all :D

• https://huggingface.co/OpenGVLab/InternVL3_5-1B
• https://huggingface.co/OpenGVLab/InternVL3_5-2B
• https://huggingface.co/OpenGVLab/InternVL3_5-4B
• https://huggingface.co/OpenGVLab/InternVL3_5-8B
• https://huggingface.co/OpenGVLab/InternVL3_5-14B
• https://huggingface.co/OpenGVLab/InternVL3_5-30B-A3B
• https://huggingface.co/OpenGVLab/InternVL3_5-38B

InternVL3_5-GPT-OSS-20B-A4B-Preview we did in the past but it only did MMPROJ extraction and InternVL3_5-241B-A28B we already successfully completed.

InternVL3_5-241B-A28B/Intern-S1 are the new winner in my private benchmark during which I manually evaluating multi-shot answers of over 100 mostly open ended mostly medical questions I'm personally interested in. They don't have a single correct answer but usually require a detailed analysis from multiple perspectives. Most of them are also not things the LLM ever saw during training as they are unique enough to not appear anywhere on the internet or in papers. I then read them all and manually judge how satisfied I'm with the answer. In my opinion an LLM winning in my benchmark gives a much better picture of a model’s true capabilities than any automated benchmark ever could.

thank you! I was wondering maybe there is some missing support in llama.cpp for some architectures

wow magically also @bartowski published these models, were they hidden or what ;)

wow magically also @bartowski published these models, were they hidden or what ;)

Yes they where hidden but not in the way you think. Basicaly when you ask llama.cpp to list all archidectures supporting MMPROJ vision stack extraction InternVLChatModel used by all thise models is missing. It took us a while to realize that its MMPROJ vision stack extraction is suported despite not beeing listed as such and add an exception to treat it as vision model to ouer system.

Yeah I made them awhile back but forgot it was private until you pinged :')

Also, I've never seen a 5.5B param vision component, those are some massive mmprojs 👀

Yeah @nicoboss discovering if vision is supported is tricky, I'm at the point where I feel like I should just run the mmproj extraction no matter what to see if they happen to work lmao

@bartowski I woudn't want that, but if it makes a significant difference, I might also consider it. Do you know of torch archs that are supported, but not marked as such? At the very least, I can trivially add the known ones to our list.