Ant's Robbyant Open-Sources Robot Vision Model
AbbreviationsEast776 · reddit · 2026-07-07
Ant Group's embodied AI subsidiary Robbyant has open-sourced the LingBot-Vision visual backbone family under the Apache-2.0 license. It includes 4 backbones ranging from 21M to 1.1B parameters, aiming to "build a brain for all robots." In contrast, Meta's DINOv3 uses a custom license with access restrictions.
In internal benchmarks, the flagship ViT-g achieved an RMSE of 0.296 on NYUv2 depth estimation, outperforming DINOv3-7B's 0.309. The distilled ViT-L essentially matches DINOv3-7B with roughly 1/23 the parameters. Its ImageNet linear probing score is 86.32 (self-tested, no independent reproduction yet), trailing DINOv3-7B's 87.87, and it also performs slightly weaker on KITTI. Note that loading requires their custom library lbotvisioninfer, rather than standard transformers or timm.
Related event: Ant Robbyant Open-Sources LingBot Vision Models, Topping Depth Benchmarks(18 posts)→
More from Models
- Local benchmark: Flint matches Qwen3.6 quality at 50/52 while running 3.6x faster on M4 Pro — RichardSocher · 2026-10-27
- ICML 2026 position paper: LLM 'thinking traces' are a harmful anthropomorphism — rao2z · 2026-10-27
- One stray 'u' and Claude starts spelling like a Brit — cantrell · 2026-10-27
- Cohere's Invent-A-Dataset paper: zero-seed synthetic data beats frontier APIs on quality and diversity — idanbeck · 2026-10-27
- ARC Prize redesigns leaderboard: highest reasoning effort by default, CSV export — mhmazur · 2026-10-27
- Cohere: open-weight Qwen3.8-Flash-Next tried uploading probe file to Beijing cloud bucket in sandbox — wstone_bd · 2026-10-27