---

base_model:

• JingyuanHuang/GUI-RD-9B

license: apache-2.0

language:

• en

pipeline_tag: image-text-to-text

library_name: transformers

tags:

• gui-grounding
• vision-language-model
• self-distillation
• gui-rd
• text-generation-inference
• qwen3.5

---

GUI-RD-9B-GGUF

> GUI-RD-9B is a 9-billion-parameter vision-language model built on Qwen3.5-9B, developed by researchers at the University of Georgia as the main experimental checkpoint for the paper "Trust the Right Teacher: Quality-Aware Self-Distillation for GUI Grounding," which addresses the task of identifying small target elements in high-resolution GUI screenshots and predicting precise screen coordinates from natural-language instructions. The model is trained using GUI-RD (GUI Reliable Distillation), a quality-aware self-distillation method that improves on naive on-policy self-distillation (OPSD) — whose coordinate-token teacher signals can become unreliable once the student-generated prefix has already deviated from the target — by combining soft correctness-aware gating, which checks whether the teacher's current coordinate-token prediction can still be completed into the ground-truth box and down-weights it if not, with teacher-probability scaling, which calibrates supervision strength using the teacher's confidence; notably, the paper found that neither mechanism improves performance alone, but combining them consistently does, since they play complementary roles. This released checkpoint corresponds to the paper's main setting (failed-token gate 0.5, scaling coefficient 3, teacher top-1 probability scaling enabled), and the method was shown to consistently improve the base model and outperform strong baselines across six GUI grounding benchmarks, making it intended for GUI grounding research and evaluation, loadable via standard Transformers classes (AutoModelForMultimodalLM/AutoProcessor) with bfloat16 precision.

Model Files

File Name | Quant Type | File Size | File Link |

|-----------|------------|-----------|-----------|

| GUI-RD-9B.BF16.gguf | BF16 | 17.9 GB | Download |

| GUI-RD-9B.F16.gguf | F16 | 17.9 GB | Download |

| GUI-RD-9B.Q2_K.gguf | Q2_K | 3.83 GB | Download |

| GUI-RD-9B.Q3_K_L.gguf | Q3_K_L | 4.93 GB | Download |

| GUI-RD-9B.Q3_K_M.gguf | Q3_K_M | 4.62 GB | Download |

| GUI-RD-9B.Q3_K_S.gguf | Q3_K_S | 4.26 GB | Download |

| GUI-RD-9B.Q4_0.gguf | Q4_0 | 5.31 GB | Download |

| GUI-RD-9B.Q4_K_M.gguf | Q4_K_M | 5.63 GB | Download |

| GUI-RD-9B.Q4_K_S.gguf | Q4_K_S | 5.35 GB | Download |

| GUI-RD-9B.Q5_0.gguf | Q5_0 | 6.31 GB | Download |

| GUI-RD-9B.Q5_K_M.gguf | Q5_K_M | 6.47 GB | Download |

| GUI-RD-9B.Q5_K_S.gguf | Q5_K_S | 6.31 GB | Download |

| GUI-RD-9B.Q6_K.gguf | Q6_K | 7.36 GB | Download |

| GUI-RD-9B.Q8_0.gguf | Q8_0 | 9.53 GB | Download |

| GUI-RD-9B.mmproj-bf16.gguf | mmproj-bf16 | 922 MB | Download |

| GUI-RD-9B.mmproj-f16.gguf | mmproj-f16 | 922 MB | Download |

| GUI-RD-9B.mmproj-q8_0.gguf | mmproj-q8_0 | 624 MB | Download |

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp