Add Qwen3.5 and a broader export path

[ilyass-afkir]

Source link

https://huggingface.co/collections/Qwen/qwen35

Target platform

iOS

Use case

Could you add Qwen3.5 (up to 4B) for on-device use? A broader Hugging Face → .aimodel export path would also be valuable, since model-by-model support does not seem sustainable as new architectures keep appearing. Thank you in advance!

Preferred precision / compression

No preference. A 4-bit option would be useful for on-device testing, with FP16 available for comparison.

Additional context

No response

[hzang001]

We do need VLMs, most importantly, (Qwen3.5) 0.8b,2b,4b

[john-rocky]

For what it's worth, in case it's useful while this stays open — there's a community zoo (not affiliated with Apple) that already has Qwen3.5 0.8B and 2B converted to Core AI .aimodel and verified on-device on an iPhone 17 Pro (iOS 27 beta), greedy top-1 exact vs the Hugging Face reference:

• Qwen3.5-0.8B — https://huggingface.co/mlboydaisuke/qwen3.5-0.8B-CoreAI (~70 tok/s on iPhone 17 Pro)
• Qwen3.5-2B — https://huggingface.co/mlboydaisuke/qwen3.5-2B-CoreAI (~28–30 tok/s)

Both are Apache-2.0 and ship int8 + fp16 bundles that run on the stock coreai-pipelined engine.

On the "broader HF → .aimodel export path" point — that's basically what the repo's conversion scripts + knowledge base try to cover (PyTorch → .aimodel convert / verify / compress, plus the gotchas for GDN-hybrid models and big-vocab heads): https://github.com/john-rocky/coreai-model-zoo (see conversion/ and knowledge/conversion-guide.md).

And @hzang001 — Qwen3-VL (2B / 4B / 8B) is in the same zoo if the multimodal angle is what you're after.

Sharing purely as a stopgap; first-party support would obviously be better.

[hzang001]

The models you shared are Confirmed by the model card — that repo are a text-only LLM, not a VLM.

[ilyass-afkir]

@john-rocky, that looks good for now. The conversion scripts are close enough to what I was looking for.