Qwen · 2026-05-28 · major
Qwen Ships Qwen-VLA — Unified Vision-Language-Action Generalist Built on Qwen3.5-4B With a 1.15B DiT Flow-Matching Action Decoder, 97.9% on LIBERO and 76.9% Real-World ALOHA
Qwen extends its 4B vision-language stack with a 1.15B DiT action decoder so one checkpoint handles robot manipulation, navigation, and trajectory prediction across embodiments. Paper, code, and demo released.
Qwen's vision-language stack now talks to robots end-to-end.
Key specs
| Backbone | Qwen3.5-4B |
|---|---|
| Action decoder | 1.15B DiT |
| Libero | 97.9% |
| Simpler widowx | 73.7% |
| Robotwin easy | 86.1% |
| Robotwin hard | 87.2% |
| R2r osr | 69.0% |
| Rxr sr | 59.6% |
| Aloha ood | 76.9% |
| Domino zero shot | 26.6% |
What is it?
Qwen-VLA is a unified embodied foundation model from Alibaba's Qwen team. It extends the Qwen3.5-4B vision-language backbone with a 1.15B DiT flow-matching action decoder, so a single checkpoint produces continuous robot actions and trajectories rather than only text. One model covers manipulation, navigation, and trajectory prediction across different robot embodiments.
How does it work?
A vision-language transformer encodes camera frames, the user's language instruction, and embodiment-aware prompt tokens describing the robot. A diffusion transformer (DiT) action decoder, trained with flow matching, then samples continuous action and trajectory tokens conditioned on those features. Joint pretraining mixes robotics manipulation trajectories, human egocentric demonstrations, synthetic simulation data, vision-and-language navigation, and auxiliary vision-language data.
Why does it matter?
Most VLA models are pinned to a single robot or task family. Qwen-VLA's single checkpoint matches or beats task-specific specialists on LIBERO (97.9%), Simpler-WidowX (73.7%), RoboTwin Easy/Hard (86.1/87.2%), real-world ALOHA OOD (76.9%), and R2R navigation (69.0% OSR), giving robotics teams a Qwen-backed starting point with paper, code, and a demo video.
Who is it for?
robotics and embodied-AI researchers
Try it
github.com/QwenLM/Qwen-VLA