Ant Group's Robbyant Open-Sources LingBot-Vision 1B Boundary-Centric Vision Model

Ant Group's Robbyant open-sources LingBot-Vision, a 1B parameter boundary-centric vision foundation model for dense spatial perception.

Robbyant, the embodied intelligence company under Ant Group, released LingBot-Vision last week. The release includes four ViT scales, with the largest at 1.1 billion parameters. The model family was trained using mask boundary modeling that treats boundaries as a native pre-training signal.

Model Family Targets Dense Spatial Tasks

The open-source package supplies ViT-g/16 at roughly 1.1 billion parameters, ViT-L at 300 million, ViT-B at 86 million, and ViT-S. All weights carry the Apache-2.0 license and sit on Hugging Face.

The 1B-scale model reaches or exceeds results from models up to seven times larger, including a 7B DINOv3 variant, on dense spatial benchmarks. Training relied on self-supervised vision transformers without labeled data.

Performance Data Shows Size Efficiency

Internal tests indicate the 1.1B model closes the gap to heavier systems on boundary-sensitive perception tasks. The approach converts boundary prediction into the core pre-training objective rather than an add-on loss.

Robbyant states the method improves spatial coherence without increasing parameter count. Independent reproduction of the numbers has not yet appeared in peer-reviewed venues.

Smaller Models Compete on Specific Metrics

Prior vision foundation models scaled parameters aggressively to improve dense prediction. LingBot-Vision instead emphasizes boundary signals during pre-training. This choice produces competitive numbers at 1B parameters against 7B baselines on the tested tasks.

The release adds concrete checkpoints across four sizes, giving practitioners options based on compute budgets. No pricing or paid tiers apply because the code base is fully open.

Open Release Adds to Growing Vision Model Options

Other labs continue to publish larger vision transformers trained on different objectives. Robbyant positions LingBot-Vision as a lighter alternative tuned specifically for boundary-aware spatial understanding. The Apache-2.0 terms allow commercial use and modification without additional restrictions.

Remaining Questions Center on Broader Validation

The initial benchmarks focus on a narrow set of dense spatial tasks. Wider evaluation across additional datasets and real-world robotics scenarios is still needed. External groups have not published independent confirmation of the 7x efficiency claim.

Watch for Follow-Up Releases and Third-Party Tests

Robbyant has not announced a timeline for new scales or fine-tuned variants. Observers will track whether subsequent papers or benchmark submissions appear in the next three months. Any third-party robotics deployment reports would supply further signal on practical gains.