# Qwen-Drive-1.0-4B Tested: One Model to See and Drive, Still Contradicts Itself > Source: > Published: 2026-09-13 00:00:00+00:00 # Qwen-Drive-1.0-4B Tested: One Model to See and Drive, Still Contradicts Itself Qwen-Drive-1.0-4B fuses perception and planning into one vision-language model. A hands-on test shows strong benchmarks but a real self-contradiction. ## What is Qwen-Drive-1.0-4B? Qwen-Drive-1.0-4B is a vision-language model from Qwen, built on top of Qwen3.5-4B, that handles both scene understanding and driving decisions using a single shared model rather than separate perception and planning systems. It adds a perception head for bird’s-eye-view detection and two planner variants (one trained by imitation, one refined with reinforcement learning) on top of the same 4-billion-parameter base, and it’s published as a single, openly downloadable repository on Hugging Face under an Apache 2.0 license. ## TL;DR - **Qwen-Drive-1.0-4B** merges perception and planning into one vision-language model built on Qwen3.5-4B, instead of the bolted-together, multi-model pipelines typical of most driving stacks. - The model adds a **perception head** that produces bird’s-eye-view detections (vehicles, pedestrians, cyclists, lane lines, drivable area) and a**planner module** that outputs a 5-second trajectory along with a natural-language explanation of its reasoning. - Two planner versions ship in the release: an **imitation-learned (SFT) planner** and a**reinforcement-learning planner** tuned for human-preference matching and closed-loop safety, with the RL version scoring higher on benchmarks like NAVSIM and the WOD end-to-end challenge. - In a hands-on local test, the model correctly described a scene as a wet road with a red light, but its own planning output said to **accelerate through a green light** , a direct contradiction of its own perception. - Loading the 4-billion-parameter model locally consumed roughly **12GB of VRAM** , making it runnable on a single consumer GPU. - The **entire release is one Hugging Face repository** with the base language model at the root and small perception/planner subfolders, rather than a sprawling multi-repo ecosystem. - Despite strong scores on driving-specific benchmarks, the model still made a basic factual error in a live test, which is the core reason it shouldn’t be trusted for real driving decisions yet. ## Seven tools to build an app. Or just Remy. Editor, preview, AI agents, deploy — all in one tab. Nothing to install. ## How does Qwen-Drive-1.0-4B differ from typical self-driving AI stacks? Most autonomous driving systems today split the job into separate components: an object detector, a mapping module, and a decision-making planner, each built and trained independently and then stitched together. That architecture makes it hard to interrogate why a car did something, because no single component holds the full picture. Ask a traditional stack why it braked, and there’s no unified reasoning trace to point to. Qwen’s approach starts from a different premise. Instead of building new specialized systems, the team took an existing vision-language model, Qwen3.5-4B, which already understands images and text jointly, and left its core architecture untouched. On top of that shared foundation, they attached two additional modules: a perception head and a planner. Both modules draw on the same underlying representation of the world that the language model already has, rather than learning to see and plan from scratch. ## What do the perception and planning modules actually do? The perception head takes multi-view camera images from the vehicle and produces a bird’s-eye-view representation of the scene. It detects vehicles, pedestrians, cyclists, cones, and barriers, and estimates drivable surfaces, sidewalks, and lane boundaries. Notably, this module was kept deliberately simple. Rather than trying to build a state-of-the-art perception system, the goal was to create a clear check on what the underlying language model itself understood about 3D space. The planner is where the model’s language ability comes into direct contact with driving decisions. Given the road ahead, it predicts a 5-second trajectory, but it doesn’t just output coordinates. It generates a short natural-language explanation first, something like deciding to decelerate to yield to a pedestrian crossing, or nudging left to pass a stopped vehicle while holding speed, and then acts on that stated reasoning. This is the central design bet: that pairing an explicit verbal justification with the trajectory output makes the model’s decisions traceable in a way bolted-together pipelines aren’t. Two planner variants exist in the release. One is trained through straightforward imitation of human driving behavior. The other adds reinforcement learning, rewarded for matching human preferences and for staying safe across closed-loop simulation. The RL-tuned planner is the one that performs better on serious planning benchmarks, including NAVSIM and the Waymo Open Dataset (WOD) end-to-end driving challenge. ## How well does it perform on benchmarks? On driving-specific question-answering benchmarks covering scene understanding, safety reasoning, and spatial awareness, Qwen-Drive-1.0-4B scores highest compared to both general-purpose vision-language models and specialized driving models built for narrower tasks. The gap isn’t marginal, it separates itself across nearly every measured category according to the reported results. What stands out is that this specialization doesn’t appear to come at the cost of general capability. When tested on unrelated images, such as math diagrams or charts with no driving content at all, the model performed at roughly the same level it did before any driving-specific training. That matters because it suggests the fine-tuning process didn’t overwrite the model’s broader visual reasoning, a common failure mode when specializing general models for narrow domains. In closed-loop simulation, where the model’s own decisions influence what happens next moment to moment, the planner has to stop at red lights, move through green ones, yield to pedestrians, and continuously adjust its path as the simulated world reacts. Trajectory visualizations show the model weighing multiple possible paths before settling on one that follows the rules of the road. ## What happened in the hands-on VRAM and accuracy test? Running the model locally involved downloading the Hugging Face repository, which bundles the base language model at the root directory alongside small dedicated folders for the perception head and both planner variants, no separate downloads or scattered dependencies required. Loading the 4-billion-parameter model and running inference pushed VRAM usage to around 12GB, putting it within reach of a single high-end consumer GPU rather than requiring data-center hardware. The test used a real driving scene: 12 camera frames captured from three different angles around the vehicle, paired with a JSON Lines file that stitches the frames together with camera angle and timestamp metadata, so the model has structured context about what it’s looking at and when. Asked to describe the scene, the model performed well, correctly identifying a wet road, a red traffic light, and a right-turn-only lane. But when asked to plan a trajectory for that same scene, the model’s stated reasoning flipped its own prior answer, saying to accelerate through a green light, despite having just correctly identified the light as red. The numerical trajectory output was consistent with that flawed reasoning, meaning the contradiction wasn’t just a wording slip, it affected the actual driving decision. ## Is Qwen-Drive-1.0-4B ready to trust for real driving decisions? Not on this evidence. The benchmark scores are genuinely strong, and the underlying idea, unifying perception and planning in one explainable model instead of a black-box pipeline, addresses a real, longstanding gap in how autonomous systems are built and audited. A model that states its reasoning in plain language before acting is a meaningful step toward interrogable AI driving systems. But the live test surfaced exactly the kind of failure that benchmarks can miss: a model that correctly perceives a red light in one output and then contradicts that perception in its very next output, acting on the wrong version of reality. That’s not a benchmark score problem, it’s a consistency problem, and consistency between what a system sees and what it decides is the entire point of a self-driving system. Strong aggregate numbers don’t rule out scene-specific failures like this one, and a single contradiction in a live test is enough to warrant caution before treating any current vision-language driving model as production-ready for real vehicles. ## Frequently Asked Questions ### What is Qwen-Drive-1.0-4B built on? It’s a fine-tuned and extended version of Qwen3.5-4B, a 4-billion-parameter vision-language model. Qwen added a perception head and two planner modules on top of the existing architecture rather than redesigning the base model. ### How much VRAM does Qwen-Drive-1.0-4B need to run? In a local hands-on test, loading and running the 4-billion-parameter model consumed approximately 12GB of VRAM, which fits on many single consumer GPUs. ### What benchmarks does Qwen-Drive-1.0-4B lead on? It scores highest among tested models on driving-specific scene understanding, safety reasoning, and spatial awareness benchmarks, and its reinforcement-learning-tuned planner performs best on planning benchmarks including NAVSIM and the WOD end-to-end challenge. ### Built like a system. Not vibe-coded. Remy manages the project — every layer architected, not stitched together at the last second. ### Does specializing the model for driving hurt its general capabilities? Based on testing described in the model’s release, no. It performed at a similar level on unrelated, non-driving images (like charts and diagrams) both before and after driving-specific training. ### Where can I get Qwen-Drive-1.0-4B? It’s available on Hugging Face as a single repository (Qwen/Qwen-Drive-1.0-4B) under an Apache 2.0 license, with the base model, perception head, and both planner variants included together.