Xiaomi OneVL - A large-scale autonomous driving model open-sourced by Xiaomi's Embodied Intelligence team.

What is Xiaomi OneVL?

Xiaomi OneVL is an open-source autonomous driving model developed by Xiaomi's Embodied Intelligence team. It is the first in the industry to unify the three major technical approaches of VLA (Vision-Language-Action) and world modeling with latent space inference into a single framework. The model uses dual auxiliary decoders to supervise compact latent variables, achieving a breakthrough where thinking is as fast as not thinking and even more effective than thinking. The vehicle-side inference latency is only 0.24 seconds, achieving state-of-the-art (SOTA) performance on four benchmarks, including NAVSIM. The model weights and code are fully open-sourced under the Apache 2.0 license.

Main functions of Xiaomi OneVL

• Visual-Language-Motion (VLA) Unified ReasoningIt integrates scene understanding, language reasoning, and driving action output into a single framework, achieving a one-step end-to-end closed loop from perception to decision-making to execution.
• World Model Future PredictionsBuilt-in visual world model decoder can predict scene frames 0.5 seconds and 1.0 seconds in the future, forcing the model to internalize the causal dynamics of road geometry, vehicle motion and environmental changes.
• Latent CoT reasoningDeep inference is performed on compact latent variables. The auxiliary decoder is discarded during inference. All latent variable tokens are pre-filled in a single parallel process with a latency of only 0.24 seconds, which is comparable to the speed of "direct prediction" but with higher accuracy.
• Dual-Auxiliary Decoder Supervised TrainingDuring the training phase, the language decoder reconstructs the textual thought chain, and the visual decoder predicts future frames. Dual-dimensional supervision ensures that latent variables simultaneously encode semantic reasoning and causal laws of the physical world.
• High-precision trajectory predictionBased on the Qwen3-VL-4B-Instruct backbone network, with an additional MLP header output trace at the top, it achieves a PDM-score of 88.84 on the NAVSIM benchmark, surpassing 8B competitors.

The technical principles of Xiaomi OneVL

• Unification of the three major technical routesXiaomi OneVL is the first to integrate three independent technical routes—VLA (Vision-Language-Action), world model, and latent space reasoning—into a single framework. This enables the model to understand the current scene while predicting future environmental evolution based on physical causal laws and output driving actions, achieving an end-to-end closed loop from perception and reasoning to decision-making.
• Dual-auxiliary decoder supervision architectureDuring the training phase, the model introduces two auxiliary decoders: a language decoder responsible for reconstructing the textual thought chain, ensuring that the latent variables encode semantic reasoning capabilities; and a visual world model decoder responsible for predicting future frames (T+0.5s, T+1.0s), forcing the latent variables to internalize physical dynamics such as road geometry and vehicle motion. During inference, both decoders are discarded, and only the compact latent variables are retained for prediction.
• Three-stage progressive trainingTraining is divided into three stages: trajectory alignment, language inference alignment, and visual world model alignment, gradually aligning latent variables with trajectory outputs, language inference targets, and visual prediction targets. Skipping any stage leads to a catastrophic performance drop (up to 21.71 points), demonstrating that joint optimization of the three stages is crucial for stable training.
• One-step latent space parallel inferenceAll latent variable tokens in the inference phase are pre-filled in a single parallel process, eliminating the need for autoregressive token-by-token generation. The latency at the vehicle end is only 0.24 seconds (4.16Hz), comparable to the speed of "directly outputting the answer." Its accuracy significantly surpasses that of the explicit thought chain method, making it the first latent space inference scheme to outperform explicit CoT in terms of performance.
• Lightweight model baseThe backbone network uses Qwen3-VL-4B-Instruct, with a compact MLP header added on top for trajectory prediction. The visual tokenizer uses Emu3.5-VisionTokenizer (codebook size 131072). With a 4B parameter scale, it achieves a PDM-score of 88.84 on the NAVSIM benchmark, surpassing 8B competitors AdaThinkDrive and LaST-VLA.

Xiaomi OneVL's core advantages

• Achieving both speed and accuracyXiaomi OneVL is the industry's first latent space inference method with accuracy surpassing explicit thought chains. Its on-vehicle inference latency is only 0.24 seconds, comparable to direct prediction speed but with better performance, and 32% faster than explicit autoregressive CoT.
• Unification of the three major technical routesXiaomi OneVL is the first to integrate three independent technical approaches—VLA, world model, and latent space reasoning—into a single framework, enabling the model to simultaneously possess scene understanding, physical causal inference, and future prediction capabilities.
• Two-dimensional explainable decisionXiaomi OneVL provides both verbal explanations and visual predictions for decision-making. It can explain driving logic in words and display future scenarios in generated images, solving the "black box" problem of unexplainable autonomous driving.
• Lightweight and high performance coexistThe model is based on the Qwen3-VL-4B-Instruct backbone network with 4B parameters, achieving a PDM-score of 88.84 on the NAVSIM benchmark, surpassing 8B competitors and achieving higher performance with a smaller model size.
• Fully open source and commercially viableThe model's weights, training, and inference code are all fully open source under the Apache 2.0 license, supporting free secondary development and commercial applications by academia and industry.
• Four benchmarks fully state-of-the-artXiaomi OneVL has achieved state-of-the-art performance in four authoritative benchmark tests: NAVSIM, ROADWork, Impromptu, and Alpamayo-R1. Among them, Alpamayo-R1 has the best ADE among all methods.

Xiaomi OneVL project address

• Project official website: https://xiaomi-embodied-intelligence.github.io/OneVL/
• GitHub repositoryhttps://github.com/xiaomi-research/onevl
• arXiv technical paper: https://arxiv.org/pdf/2604.18486

Xiaomi OneVL Competitive Product Comparison

Application scenarios of Xiaomi OneVL

• Advanced intelligent driving mass production deploymentIt addresses the bottlenecks of traditional CoT inference, such as excessive latency and inability to deploy in real time.
• Decision making in complex road conditions: Predict pedestrian intentions, vehicle lane changes and other dynamic scenarios to avoid "black box" risks.
• Model Decision VisualizationIt provides auditable and interpretable language and vision-based decision-making support for autonomous driving systems.
• Secondary development of academic researchOpen source code and weighted support for commercial use by academia and industry under the Apache 2.0 license.