The Pulse
Ant Group Releases GUI-Trained LBS-27B Vision Model
Ant Group’s inclusionAI released LBS-27B, a vision-language model adapted using synthetic GUI-grounding data, on October 26. The team reports results on five visual benchmarks and publishes the weights under Apache 2.0, while noting that li
AI.info Team · Written with AI · fact-checked ·
89.13% on HR-Bench 4K
Ant Group’s inclusionAI says its LBS-27B model scored 89.13% on HR-Bench 4K, a benchmark for answering questions about high-resolution images. The model, released on October 26, also scored 87.75% on HR-Bench 8K and 81.15% on ScreenSpot-Pro, which measures accuracy in locating interface elements. Its weights and project software are published under the Apache 2.0 license.
LBS-27B is based on Qwen3.8-27b and is designed to inspect images by selecting regions, zooming in and revisiting earlier crops. inclusionAI describes the approach as transferring a visual-observation policy learned through graphical user interface (GUI) grounding to general visual question answering. The training claim applies to the data used for this adaptation; the model card does not say the underlying Qwen base model was trained only on GUI data.
“It learns to select and inspect image regions through GUI grounding, then transfers this observation policy to general-domain visual question answering and high-resolution perception.”
inclusionAI, LBS-27B model card
Training stays within synthetic GUI data
The team says it used approximately 10,000 supervised fine-tuning examples for a cold start, then trained with 80 steps of Group Relative Policy Optimization, or GRPO. Each step used a batch size of 64 and eight rollouts per prompt. The model card says those training stages used synthetic GUI-grounding data, without general-domain data or visual question-answering examples.
During use, LBS can request as many as two separate crops in a turn, with a limit of three crop calls per episode. It can refine a crop or return to the original image, an approach intended to help it locate small interface targets and examine detail in high-resolution images. The released checkpoint is available for serving with a compatible vLLM setup; inclusionAI’s example uses four devices.
The scores come with comparison limits
The results cover five evaluations: ScreenSpot-Pro, OS-World-G (Refine), PerceptionBench, and HR-Bench at 4K and 8K. On OS-World-G, the reported 84.57% action accuracy includes refusal cases. PerceptionBench’s 59.67% judge accuracy comes from 3,000 records scored by a self-hosted Deepseekv4-flash-0731 judge.
Those figures are not a controlled head-to-head comparison with other models, the release notes say. The reference results use different inference setups, and the PerceptionBench comparisons use a different judge, GPT-oss-120B; repeated judgments may also vary. The score table therefore records inclusionAI’s reported results, not proof that LBS-27B outperforms a named competitor under identical conditions.
Coordinates do not operate an app by themselves
LBS returns plain-text crop requests and normalized coordinates for interface targets. A client must carry out the crop requests, return the images and coordinate mappings, and—if the goal is to operate a live application—send the resulting action to a separate executor. The model locates a target; it does not itself click through an app.
The weights are downloadable, but third-party datasets and screenshots remain subject to their original licenses and terms. The concrete test ahead is whether the GUI-trained observation policy transfers reliably to visual questions outside interfaces when independent teams evaluate it with matched benchmarks and scoring protocols.