Qwen-Drive-1.0-4B

Qwen-Drive-1.0-4B

A driving model built on Qwen3.5-4B that sees in 3D, answers questions about road scenes and plans future vehicle paths. For researchers working on autonomous driving.

Made by Qwen on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Qwen-Drive-1.0-4B is a driving-focused vision-language model from the Qwen team. It starts from the Qwen3.5-4B model and leaves that model's design untouched. Two add-on parts sit beside it. One is a perception head that finds 3D objects, predicts occupied space and segments a bird's-eye-view map. The other is a Planning Expert that turns the model's understanding of a scene into a path the vehicle could follow. The same model also answers open questions about a driving scene in plain text. This repository holds the weights and configuration files only.

Who it suits: Researchers and engineers studying autonomous driving who want one model for scene questions, 3D perception and motion planning.

Industry | Who uses it | How they use it | Example
Autonomous driving | Perception researchers | Test 3D detection and map output from one model | Inspect the bird's-eye-view output for a scene
Autonomous driving | Planning researchers | Compare planner variants on driving benchmarks | Run the direct and the reasoning planning modes
AI research | Vision-language teams | Study how driving data affects a general model | Ask free-form questions about a road image

Anyone who wants a ready driving app can skip it, since the card sends the code, demo data and documentation to a separate project.

Where to use it:

• The transformers library in Python, with the Safetensors weights.
• The Qwen-Drive-1.0 project on GitHub, which holds the matching code and demo data.
• Windows, macOS or Linux: the download steps run on all three with Python.

Use case | Who | What they get back
Scene questions | Researchers | Text answers about a driving image
3D perception | Perception engineers | Detected objects, occupancy and a map segmentation
Motion planning | Planning researchers | Future paths for the ego vehicle

How to set it up: What you need first:

• Python with pip.
• The transformers library.
• A GPU with enough memory for the weights, about 9.1 GB in BF16 or about 2.3 GB at 4-bit, as estimated on the record.
• The code from the project on GitHub, for perception and planning.

1. Install the Hugging Face hub package with pip.
2. Download the repository with the hf command.
3. Fetch the project code and demo data from GitHub.
4. Follow the model page for the exact way to load and run it.

pip install -U huggingface_hub
hf download Qwen/Qwen-Drive-1.0-4B

How to use it: You can ask the base model free-form questions about a driving image. For perception and planning you also need the attached modules from the project code. The text available here leaves out the run steps, so use the model page and the GitHub project for the calls.

Title: Parts of the model
Tool | What it does
Qwen3.5-4B backbone | Reads the image and answers questions in text
BEV perception head | Finds 3D objects, occupancy and map areas
Planning Expert | Generates future paths for the ego vehicle
planner-sft | Supports direct and reasoning planning
planner-rl | Reward-tuned planner that works best in reasoning mode

Example: You give the model a front camera frame and ask what is ahead. It answers in text. With the planner attached, the same scene also produces a predicted path for the vehicle.

For newcomers and experts: Think of it as a language model that learned to drive. It can talk about a road scene, point out objects in 3D and propose where to go next. A good first step is to ask the base model a few questions about a single driving image before trying the planner.

The planner uses flow matching to generate paths, and it was tuned with rewards from NAVSIM and WOD-E2E scoring plus a displacement term. The perception head is kept outside the VLM on purpose, so it works as an inspectable 3D probe. Training mixes driving data with general vision-language data to limit forgetting.

Read a driving scene as an image
Encode it with the Qwen3.5 backbone
Run the perception head or the planner
Get 3D output or a future path

Good to know: The record estimates how much space the weights take.

Title: Weights size, estimate
BF16: 9.1 GB
4-bit: 2.3 GB

• Size: about 4.5B parameters.
• Versions: the record lists 4 quantized and 8 fine-tuned versions made by others.
• Training: driving data is mixed with general vision-language data.
• License: apache-2.0
• Cost to run: no API key, account or paid service named.
• Data it reaches: files you download from Hugging Face and the images you pass in.
• Not for: anyone who needs a finished driving app, or a repository that includes the code and demo data.

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 4.5B
• Weights: about 9.1 GB in BF16; about 2.3 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 7,740
• Likes: 259
• License: apache-2.0
• Based on: Qwen/Qwen3.5-4B
• Quantized versions: 4
• Fine-tunes: 9
• Tags: transformers, safetensors, qwen_drive, autonomous-driving, motion-planning, 3d-perception, visual-question-answering, qwen, image-text-to-text, conversational

Links

• Model: https://huggingface.co/Qwen/Qwen-Drive-1.0-4B
• Files: https://huggingface.co/Qwen/Qwen-Drive-1.0-4B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Qwen/Qwen-Drive-1.0-4B

Then follow the usage in its model card. Nothing in this kit runs by itself.