Alibaba unveils its Qwen Robot Suite for autonomous robots

Alibaba Unveils Its Qwen Robot Suite for Autonomous Robots

Alibaba has just introduced the Qwen Robot Suite, its first suite of AI models designed for autonomous robots. Tongyi Lab, the group's research division, has developed three distinct components, including a video model of the physical world that currently has no competitors. The Chinese firm has begun testing with a select group of enterprise clients from Alibaba Cloud.

Tongyi Lab has divided the robot's intelligence into three layers, each assigned to a distinct model.

• For navigation, the lab developed Qwen-RobotNav, a spatial analysis and movement calculation model.

• Qwen-RobotWorld was designed to predict the evolution of a physical scene before any mechanical action. This is the video model of the suite.

• For physical execution, the lab developed Qwen-RobotManip on the Qwen3.5-4B architecture, a model intended for robotic arms.

DAMO Academy, Alibaba's research division, released RynnBrain as open source on GitHub as early as February. This model was built on Qwen3-VL to provide robots with spatial perception grounded in the physical world. Alibaba has initiated testing with a select group of enterprise clients from Alibaba Cloud, but little information has been disclosed regarding the names or methodology.

Qwen-RobotManip and Its Performance

Among the three models, Qwen-RobotManip has received the largest dataset. Tongyi Lab injected over 38,000 hours of training data, including subjective view captures and open-source games derived from real robots. On the RoboChallenge Table30 v1 benchmark, in the generalist category, Alibaba claims the top spot for Qwen-RobotManip, with a success rate of 45% and a process score of 59.83, which is 20% above the second place.

DAMO Academy distributes RynnBrain in four variants:

• Two dense versions with 2 and 8 billion parameters.

• A mixture-of-experts version with 30 billion parameters.

• Specialized models by use case: planning, navigation, and spatial reasoning.

During a demonstration by DAMO Academy, a robot was able to identify a fruit and place it in a basket.

Innovations from Alibaba Cloud

At the Alibaba Cloud summit in Hangzhou in May, Liu Weiguang, Senior Vice President of Alibaba Cloud, stated that Alibaba alone covers the entire AI stack in China, from chips to agent applications.

Qwen3.7-Max for Autonomous Agents

For the reasoning layer, Alibaba has chosen Qwen3.7-Max, launched on May 20. Unlike previous open-source Qwen models, Qwen3.7-Max is only accessible via an API, priced at $2.50 (2.24 euros) per million tokens in input and $7.50 (6.72 euros) per million tokens in output, through Alibaba Cloud Model Studio. This model was designed for long tasks in production conditions.

During an internal stress test on a novel hardware platform, Qwen3.7-Max operated for 35 hours without interruption, with 1,158 tool calls and 432 kernel evaluations. At the end of this cycle, the inference speed of the Zhenwu M890 processor, Alibaba's latest AI accelerator, had improved by a factor of ten. Alibaba produced these figures internally, without verification by an independent body at this stage.

On the Apex Math benchmark (mathematical reasoning), Alibaba claims 44.5 points for Qwen3.7-Max, compared to 34.5 for Claude Opus 4.6 Max and 38.3 for DeepSeek V4-Pro Max. Additionally, developers can integrate it into environments like Claude Code or Qwen Code via the MCP protocol, without specific reconfiguration.

In January, developers had downloaded the Qwen models 700 million times on Hugging Face. Third parties had derived 180,000 versions in 119 languages. Tongyi Lab plans to release the complete suite in June 2026.