AI/TLDR

Qwen-Drive-1.0-4B

Published August 31, 2026: an Apache-2.0 4B vision-language foundation model for autonomous driving that unifies 3D perception, driving VQA and motion planning on an unchanged Qwen3.5-4B backbone.

Qwen-Drive (autonomous driving VLM)Open weightsAvailable — open weights on Hugging Face
Released
31 Aug 2026
Parameters
4B (Qwen3.5-4B backbone, 9.1 GB), plus a 0.5 GB BEV perception head and a 2.1 GB planning expert
License
Apache-2.0

Overview

Qwen-Drive-1.0-4B is Qwen's first vision-language foundation model aimed at autonomous driving, published with a technical report dated August 31, 2026 and titled "Qwen-Drive-1.0: An Initial Step towards a Vision-Language Foundation Model for Autonomous Driving". The point of the release is unification: 3D perception, driving visual question answering and motion planning are handled by one model rather than by three separate stacks bolted together.

It is built on Qwen3.5-4B, and the model card is explicit that the pretrained architecture is left entirely unchanged. Driving capability is added around it — a BEV perception head that does 3D detection, semantic occupancy and map segmentation, and a planning expert that emits trajectories. That separation is deliberate: the perception head doubles as an explicit 3D probe of what the language model's own representations already encode.

The planning expert comes in two variants. `planner-sft` is supervised and supports both direct and reasoning-based planning; `planner-rl` is further trained with reinforcement learning against NAVSIM PDMS, Waymo Open Dataset E2E and displacement metrics, and is the one the card recommends for reasoning-based planning. Inference returns both the reasoning trace and a set of sampled trajectories — the demo produces six candidates covering five seconds at 10 Hz.

Qwen reports results in three families. On planning: 90.7 NAVSIM PDMS in the pseudo-closed-loop setting (91.4 best-of-6), 8.45/7.91 WOD-E2E RFS on val/test, 1.27/2.67 m WOD-E2E ADE at 5s, 0.42 m average ADE at 3s on PAI-AV, and an AlpaSim at-fault score of 0.37 closed-loop. On driving VQA: 77.8 on LingoQA, 74.5 on WaymoQA (70.7 on its safety split), 66.5 on VLAD, 66.1 on SURDS and an Ego3D RMSE of 7.78 m.

The third family is the one that makes it a foundation model rather than a driving head: general vision-language scores are reported alongside the driving ones — 85.5 MMBench, 75.9 MMStar, 72.7 MMMU, 62.7 MMMU-Pro (standard), 64.4 CharXiv, plus 78.9 EmbSpatial, 48.5 ERQA and 45.8 Omni3D for spatial understanding. Qwen attributes this to staged training that mixes driving supervision with general vision-language data specifically to avoid catastrophic forgetting.

Weights are on Hugging Face under Apache-2.0, with the code in QwenLM/Qwen-Drive-1.0 and a demo script carrying four planning scenes and six perception frames.

Released2026-08-31
LicenseApache-2.0
WeightsOpen weights
Parameters4B (Qwen3.5-4B backbone, 9.1 GB), plus a 0.5 GB BEV perception head and a 2.1 GB planning expert
ArchitectureVision-language model with two external heads. The Qwen3.5-4B encoder is kept architecturally unchanged and answers free-form questions about the scene; a BEV perception head jointly does 3D object detection, semantic occupancy prediction and BEV map segmentation as an explicit 3D probe of the VLM's representations; a planning expert generates trajectories, shipped as `planner-sft` (direct and reasoning-based planning) and `planner-rl` (further optimised with reinforcement learning against NAVSIM PDMS, Waymo Open Dataset E2E and displacement metrics).
ModalitiesText, Image
StatusAvailable — open weights on Hugging Face

Benchmarks

  1. NAVSIM PDMS (pseudo-closed-loop)90.7
  2. LingoQA (driving VQA)77.8%
  3. WaymoQA (all)74.5%
  4. MMBench85.5%
  5. MMMU72.7%
  6. EmbSpatial78.9%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

Strengths

  • One model covers 3D perception, driving VQA and motion planning instead of three separate stacks
  • 90.7 NAVSIM PDMS (91.4 best-of-6) with an RL-trained planning expert
  • Retains general vision-language ability — 85.5 MMBench and 72.7 MMMU alongside the driving scores
  • The Qwen3.5-4B backbone is unchanged, so the driving capability rides on a standard VLM rather than a bespoke architecture
  • Apache-2.0 weights at 4B, small enough to run and fine-tune outside a large cluster
  • Reasoning-based planning returns the trace alongside the sampled trajectories

Best for

  • Research on end-to-end driving policies that need perception, language and planning in one model
  • Probing what a vision-language model's representations encode in 3D, via the BEV perception head
  • Generating and scoring candidate trajectories in open-loop, pseudo-closed-loop and closed-loop evaluation
  • Driving-scene question answering where an explanation matters as much as the trajectory
  • A permissively licensed 4B starting point for domain fine-tuning on proprietary driving data

FAQ

Is Qwen-Drive-1.0-4B open weights?

Yes. The weights are published on Hugging Face under Apache-2.0, together with the code at github.com/QwenLM/Qwen-Drive-1.0.

What is it built on?

Qwen3.5-4B. The model card states the pretrained architecture is kept entirely unchanged; the driving capability comes from an external BEV perception head and a separate planning expert.

What is the difference between planner-sft and planner-rl?

planner-sft is the supervised variant and supports both direct and reasoning-based planning. planner-rl is further optimised with reinforcement learning against NAVSIM PDMS, Waymo Open Dataset E2E and displacement metrics, and is what the card recommends for reasoning-based planning.

Does it still work as a general vision-language model?

Qwen reports general benchmarks alongside the driving ones — 85.5 MMBench, 75.9 MMStar, 72.7 MMMU and 62.7 MMMU-Pro (standard) — and says staged training mixes driving supervision with general vision-language data to limit catastrophic forgetting.

What does inference return?

In reasoning-planning mode the model returns a reasoning trace plus a batch of sampled trajectories; the demo requests six samples and returns a (6, 50, 3) tensor — five seconds at 10 Hz.