Distinct repos for each checkpoint
Hello!
Firstly, congrats on being one of the top 100 trending models of the day. To increase visibility and usage I would recommend two things:
- For each checkpoint you should have a different repo. For ex.
turboderp/Qwen3-Next-80B-A3B-Instruct-exl3-2.08bpwand so on. We can later add all of them to a collection namedQwen3-Next-80B-A3B-Instruct-exl3(see more collections: hf.co/collections) - Each repo should have
library: transformersin the readme metadata, which would let people know that they can run inference using the transformers library.
If you need any help from my side, let me know. Looking forward to hearing from you.
EXL3 models don't directly load in Transformers, though, not without an experimental integration option that's mostly for development work and not all that performant. And for this model in particular, with its about 220k individual tensors, Transformers can't really handle it anyway and just gets stuck in some O(n^2) bookkeeping logic.
As for the checkpoints, isn't it better to use revisions for different quantization levels of the same model? If I split every quantized model into ~10 different repos, that's a lot of new repos I'd have to create. Search results are kind of cluttered already.
experimental integration option that's mostly for development work and not all that performant.
FYI - I love this hidden / unexpected feature (can't even remember how I stumbled upon it) and have been using it a lot (patiently) with Qwen3-235B-A22B-Instruct-2507.
So thanks for that.