GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
GUIAgentt
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
GUIAgentt
Model Tree
Qwen/Qwen3.5-9B
Input Modalities
Output Modalities
Supported Functionality
The model takes a screenshot and GUI interaction history, reasons inside
<think>...</think>, and emits a structured <tool_call> action. It is intended
for research and evaluation of Korean-language computer-use agents.
The included chat_template.jinja defines the GUI18 conversation and tool-call
format used during training. Load the processor from this repository so that the
same template is retained.
Qwen/Qwen3.5-9BThe adapter was merged with PEFT safe merge and the resulting standalone model was successfully reloaded for validation. Both language and vision LoRA modules were trained.
python
import torchfrom transformers import AutoModelForImageTextToText, AutoProcessor model_id = "GUIAgentt/Qwen3.5-CUA-Ko100-human" processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)model = AutoModelForImageTextToText.from_pretrained( model_id, dtype=torch.bfloat16, device_map="auto", trust_remote_code=True,)Use the processor's bundled chat template when rendering multimodal messages and tool definitions.
The model was optimized for the training GUI schema and action vocabulary. It may not generalize to other tool schemas without prompt or template adaptation. The validation split covers the ARM4 portion of the training mixture, so it does not independently measure every data source. Model outputs should be validated before deployment in environments where actions can modify user data.
Apache License 2.0, following the base model.