Alibaba’s Qwen team releases open-source model for autonomous driving

Summary

Alibaba's Qwen team, in collaboration with Huazhong University of Science and Technology, released Qwen-Drive-1.0-4B, an open-source autonomous driving model that integrates 3B perception and trajectory planning modules onto the Qwen3.5-4B vision-language backbone. The project includes code, weights, and demo data under the Apache 2.0 license; it also offers two planning variants, one optimized via imitation learning and the other via reinforcement learning.

Why it matters

By bringing a VLM-based driving stack to the open-source community, this release lowers the barrier for academic labs and smaller automakers to conduct end-to-end driving experiments without relying on the closed stacks of Tesla or Waymo. This step also marks Alibaba's concrete entry into the embodied, safety-critical AI domain, extending beyond its text-generation LLMs, a field where open-source tools are notably scarce. The Apache 2.0 license enables commercial use, potentially accelerating the formation of a broader ecosystem around China-originated driving models.

Highlights

• Built on the Qwen3.5-4B VLM with separate 3B perception and trajectory generation modules
• Two planning heads optimized with imitation learning and reinforcement learning, sharing the same VLM backbone
• Released under the Apache 2.0 license, including code, model weights, and demo data

Critical take

The paper is surprisingly weak: it lacks any benchmark data, fails to compare against existing open-source driving references (such as UniAD or VAD), and does not address hardware requirements or real-world closed-loop testing. The fact that a 4B-parameter VLM handles full driving perception and planning raises legitimate questions about whether the system can generalize beyond curated demo data to the long-tail complexity of real-world road scenarios.