Gasgoo Munich- On May 13, Xiaomi introduced Xiaomi OneVL, a new AI framework designed for autonomous driving that combines latent-space reasoning with language and visual intelligence in a single-step architecture.
Image source: Xiaomi
According to the company, the system brings together several major technical approaches — including vision-language-action (VLA) models, world models, and latent-space reasoning — within one unified framework. Xiaomi said the architecture is capable of delivering stronger reasoning performance while also improving both inference speed and accuracy compared with existing methods.
Xiaomi said OneVL adopts a dual-supervision mechanism that combines language-based reasoning with future visual prediction. The approach is intended to integrate explainability and predictive world-model capabilities directly into the latent reasoning process, allowing the model to better understand and anticipate complex driving environments.
The company explained that the framework is built around the idea that autonomous driving requires more than compressed language reasoning alone. Instead, effective decision-making depends on a broader understanding of how the visual world is likely to evolve over time.
In practical driving scenarios, factors such as vehicle motion, road geometry, and the dynamic behavior of surrounding obstacles are governed by complex spatial and temporal relationships. Xiaomi argues that relying solely on compressed language representations risks losing critical structural information, while compressing predictions of future visual scenes can preserve the elements most relevant to driving outcomes.
Building on that concept, Xiaomi developed three core technologies aimed at enabling the model to "think" using its own internal representation system, predict future visual scenarios, and compress the entire reasoning pipeline into a single inference step.
The company said Xiaomi OneVL achieved record-setting results across several mainstream reasoning and planning benchmarks for autonomous driving applications. According to Xiaomi, the framework outperformed existing latent-space reasoning approaches in accuracy, surpassing explicit chain-of-thought (CoT) reasoning methods while maintaining inference speeds comparable to lightweight "answer-only" prediction models.
Lei Jun, Xiaomi's founder, CEO, and chairman, said Xiaomi plans to fully open-source both the model and its underlying codebase, inviting developers and researchers worldwide to participate in advancing large-scale AI models for autonomous driving.
