MiDashengLM-Spatial
Xiaomi open-sourced MiDashengLM-Spatial, an audio-language model that understands what a sound is and where it comes from, with weights and code under Apache 2.0.
It is built on MiDashengLM-7B and adds a spatial audio encoder called Spatial-Dasheng. A hierarchical semantic-to-spatial conditioning module captures semantic and spatial information separately and then merges them. The model reasons about direction, distance and movement of sound sources from binaural audio, and keeps speech recognition, captioning and music and environmental sound question answering.
- Date
- Friday 16 October 2026
- Lab
- Xiaomi
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Parameters | about 8 billion as listed on the Hugging Face page, per Pandaily | company |
Xiaomi calls it the industry's first open-source end-to-end spatial audio understanding model, a claim not independently assessed. It was trained on binaural audio, and mono input runs without directional information.
Sources
- pandaily.com/xiaomi-midashenglm-spatial-open-source-spatial-audio-understanding-model-apac
- technode.com/2026/10/16/xiaomi-open-sources-audio-ai-model-that-can-identify-where-sounds-
This record was checked against its sources on 16 October 2026. How we check