Comment by 3eb7988a1663
5 days ago
If there were ever a sign I need to get mirroring, this is it.
What are the best Chinese models on HuggingFace today? Bucket by ideal RAM: <16GB, <32GB, <96, <256, 256+
Text generation, image generation, TTS, etc
5 days ago
If there were ever a sign I need to get mirroring, this is it.
What are the best Chinese models on HuggingFace today? Bucket by ideal RAM: <16GB, <32GB, <96, <256, 256+
Text generation, image generation, TTS, etc
256+: GLM-5.2 (swap with Kimi K3 when it goes open-weight on Monday)
<256: Actually, surprisingly, not a Chinese model but probably Laguna S2.1. The best Chinese model at this size is DeepSeek V4 Flash though
<96: Qwen 3.6 27B
<32: Still Qwen 3.6 27B (NVFP4)
<16: Oof, not sure. Nothing will feel great at this size TBQH without finetuning on a specific task. Pick your poison of tiny Qwen or tiny Gemma (although again Gemma is not Chinese)
Exactly the kind of breakdown I was hoping to see. Thanks.
And we can quickly see that the real problem is not the models, but HW to run them. You can build whole Enterprise on Gemma 4 31B full precision without significant problems. If you can afford not to lobotomise it by quantisation.
[dead]
Alibaba Cloud runs ModelScope, which is like Hugging Face but based in China:
- https://modelscope.ai
- Source code (ironically on GitHub): https://github.com/modelscope/modelscope
It's not limited to Alibaba's Qwen. All of the other major Chinese ones are on there (GLM, DeepSeek, Kimi, etc.) as well as finetunes and quantizations of the non-Chinese ones.
The two major image generation models are both by Alibaba - Z-Image and Qwen-Image-2512/Qwen-Edit-2511. They are both decent but have different use-cases. Unfortunately with the recent Qwen-Image 2 and now 3 it doesn't seem likely that they're going to be doing any more open-weight image gen.
https://huggingface.co/Tongyi-MAI/Z-Image-Turbo
https://huggingface.co/Qwen/Qwen-Image-2512
https://huggingface.co/Qwen/Qwen-Image-Edit-2511
My list would be to go grab GLM5.2, Deepseek v4 Flash and Pro, and the various sizes of Qwen 3.5/3.6 dense and MoE, along with at least one of the multi modal qwens (I think it’s called VL). In a pinch, you can quantize them yourself to whatever size you need.
"you can quantize them yourself" ... to the point where usual Excel will have more intelligence.
Hugging Face recently added a filter directly on the models page: https://huggingface.co/models that lets you filter based on hardware.
Filtering based on hardware is super helpful. Everyone should do this. There are millions of models, and most of them require a server room to load and run.
Specific use case is way too important for a list like that to be very helpful. I only know image gen well enough to use as an example, but style, resolution, inpainting, img2img or text2image, lora compatibility, etc. all lead to different “best” models.
Even more complex for image2video, but there’s fewer models to choose from there at least.
Sanctions would cancel the whole business rationale for the Chinese government having/supporting open weight models. You can't commoditize your competitors with open-source if your product is blocked.
Even with US sanctions, Chinese AI companies would continue to serve users in every other country in the world (including China itself), so I don't see why they would stop releasing open weight models.
The open weights models would be even more effective by incentivizing people and companies to switch jurisdictions. It might even break SFs monopoly on AI, or at least weaken it, everybody isn’t there exclusively to funnel money to openAI and Anthropic.
It would also boost research in non US jurisdiction. Who is going to be wooed by “come to our lab where you’re only allowed to work with closed models!”
5 replies →
Open-weights models will be downloaded from China and run on local/corporate-owned AI hardware. You would have to ban Internet access and AI hardware next.
1 reply →
It is a lost opportunity that huggingface (or unsloth) frontpage doesn't have a list like this that updates which each model release.
From the HF front page if you hit "models" it should auto sort by trending before you change any search parameters, so if anything large, capable and popular has just been released it's almost certainly going to be at the top of the list.
Hugging Face recently added a filter directly on the models page: https://huggingface.co/models that lets you filter based on hardware.
qwen3.6-27B, gemma4-it-31B, poolside/Laguna-S-2.1 are all good text gen models generally in the <128GB (depending on quant, you can fit them in smaller mem footprint)
Note, Laguna provided quants have some issues and they are reworking/updating them.