Has anyone tried the quantized model for this

Open codebanesr opened this issue 8 months ago • 2 comments

I see two quantized versions for 1.5 on huggingface https://huggingface.co/models?other=base_model:quantized:ByteDance-Seed/UI-TARS-1.5-7B

Does this also suffer from performance degradation like v1 when quantized ? Also, when can we have a managed model deployed on modelscope ?

Apr 23 '25 20:04 codebanesr

I've tried several quantized model of UI-TARS-7b-DPO, the grounding performance can be affected since the model always predicts the wrong coordicates

Apr 27 '25 04:04 ZeruiQin

We do not recommend using the quantized versions of the model, as they have not undergone rigorous performance validation.

Apr 29 '25 15:04 JjjFangg

refer: https://modelscope.cn/models/HHaoAI/UI-TARS-1.5-7B-GGUF

only support LM Studio.

May 09 '25 02:05 hhao