Xiaomi has released MiDashengLM-Spatial, a new member of its Dasheng audio model family that is designed to understand not only what a sound is, but also where it comes from. The company published the full model weights and inference code under the Apache 2.0 license, which permits both research and commercial use.
The model is built on MiDashengLM-7B, the sound-understanding large model Xiaomi open-sourced earlier, and adds a spatial audio encoder called Spatial-Dasheng. According to the technical report posted to arXiv, a hierarchical semantic-to-spatial conditioning module captures semantic and spatial information separately and then merges them. The result is a single end-to-end model that can describe speech, environmental sounds and music while also reasoning about a source's direction, distance and movement. The Hugging Face listing puts the model at roughly 8 billion parameters.
Spatial cues depend on multi-channel input. The model card says the model was trained on binaural, two-channel audio and that mono input still runs but carries no directional information. Xiaomi says the model keeps its general audio abilities, and the research team reports that performance on basic audio tasks remains comparable to large audio-language models of around 8 billion parameters.
Xiaomi describes MiDashengLM-Spatial as the industry's first open-source end-to-end spatial audio understanding model. The paper makes a similar claim, qualified with the phrase "to our knowledge." That claim comes from Xiaomi and its researchers and has not been independently assessed.
Over the past two years Xiaomi has open-sourced the Xiaomi Dasheng sound foundation model and MiDashengLM-7B. The company says technology from the Dasheng series already powers more than 30 applications across its smart home and in-car cockpit products, covering speech, ambient sound and music perception. That figure refers to the existing series, not to the new model.
For MiDashengLM-Spatial itself, Xiaomi says it will be rolled out gradually in smart home and car cockpit scenarios, where knowing where a voice or alarm is coming from could make assistants respond more precisely. The company has not given a timeline or named specific products.
The release includes a project demo page, the arXiv technical report, evaluation code on GitHub and model weights on Hugging Face.
SEE ALSO: Xiaomi Open-Sources MiMo-V2.6 Pro and Flash Weights Plus RL Stack