Xiaomi has opened a model that asks audio AI to do more than recognize a sound or turn speech into text. MiDashengLM-Spatial is designed to retain the directional clues in binaural recordings, so a system can reason about whether a sound is to a listener’s left or right, how far away it is, whether it is moving, and how it relates to other sounds. That distinction changes the question for developers building voice interfaces: recognizing a door closing is useful; locating the door relative to a user is a different capability.
TechNode reported that Xiaomi has open-sourced the model, its weights, and inference code under Apache 2.0. The public release is also unusually tangible for an audio-model announcement: Xiaomi’s Research GitHub repository includes installation, inference, and evaluation instructions. For engineers, that means the central claim can be inspected through code and a repeatable test path rather than treated as a demonstration alone.
An Open Xiaomi Model Built to Keep Direction Intact
The work comes from Xiaomi’s Xiaoai Plus Dasheng Team, identified in the associated paper as part of Xiaomi Inc. That connection matters because MiDashengLM-Spatial extends Xiaomi’s Dasheng and MiDashengLM audio research line, rather than representing another release in the company’s separate MiMo language-model work. Xiaomi’s earlier public work on MiMo-V2.6 reinforcement learning concerned a different problem: improving agents’ language-model behavior. This release centers on what microphones can tell a model about a physical scene.
The package spans both software and model files. Xiaomi’s code repository carries the Apache License, Version 2.0, while its Hugging Face model repository lists seven safetensors weight shards alongside configuration, processor, tokenizer, and model-code files. The displayed file total is 33.5 GB, a practical indication that this is a substantial download for a developer exploring the released checkpoint, not a statement about runtime memory or deployment cost.
Apache 2.0 makes the code and model available as Xiaomi presents them for research and commercial use. The license should not be read as a blanket answer about training data, other dependencies, trademarks, or every downstream use. The useful point for builders is narrower: they have public materials with which to inspect the model and run the company’s supplied workflow.
At the center of that workflow is an input constraint that is easy to miss in a headline. Xiaomi’s instructions call for real binaural audio in a two-channel (2, T) arrangement, placed before the text prompt. Mono audio can still run through the model, but Xiaomi says it carries no directional information. A spatial answer therefore depends not only on the model but on the microphone arrangement and the audio pipeline delivering the two channels.
Binaural Input Turns Sound Recognition Into Spatial Reasoning
Xiaomi describes MiDashengLM-Spatial as a combination of MiDashengLM and a Spatial-Dasheng encoder, joined through what it calls hierarchical semantic-to-spatial conditioning. In plain terms, the design is meant to connect an understanding of what the sound is with clues about where it is and how it changes. Xiaomi’s repository and paper name sound-event localization and detection, direction of arrival, source distance, source motion and trajectory, spatial audio captioning, spatial speech transcription, and spatial question answering among the tasks it supports or evaluates.
The model also retains general-audio tasks, including automatic speech recognition, audio captioning, audio question answering, and questions about environmental sound and music. That breadth is why the release is potentially more interesting than a narrowly tuned localization component. A product team would not necessarily want one system for transcription and another for a sound source’s direction. Still, the materials establish supported and evaluated tasks, not reliable behavior in a consumer product.
The training description underscores the scale of the spatial problem Xiaomi is trying to model. Its paper says the team generated one million synthetic spatial acoustic scenes, roughly 13,000 hours of binaural audio, and six million fact-verified question-and-answer pairs. The clips ran from 30 to 60 seconds and involved no more than three overlapping sources, using simulated room impulse responses and head-related transfer functions. Those figures describe Xiaomi’s disclosed methodology, not an independently audited dataset.
That setup also illuminates the reader stake. A home assistant or in-car interface needs more than a transcript if it is expected to distinguish a spoken request from a sound on one side of a room or car. TechNode reported that Xiaomi sees possible use in its home-device and automotive ecosystems, without identifying a shipping product, timetable, integration partner, or production deployment. The immediate value is a public research and development path, not a new Xiaomi feature for buyers.
Xiaomi’s Benchmark Results Put Channel Swaps at the Center
The most revealing part of Xiaomi’s evaluation is not a generic audio score. It is whether an answer changes when left and right channels change. The repository supplies one-shot evaluation scripts for the spatial-reasoning subsets of STAR-Bench and MMAU-Pro, allowing developers to examine tests that probe directionality rather than simply ask whether a model heard an event.
In Xiaomi’s evaluation, MiDashengLM-Spatial-7B recorded 70.3 spatial accuracy on the STAR-Bench Spatial Reasoning subset. Xiaomi also reported a 64.1 option-swap all-correct rate and a 72.6 channel-swap all-correct rate, tests designed to expose whether the model follows positional cues. Its table lists a 73.7 human spatial-accuracy reference and 47.0 for Gemini-3.6-Flash, but those are Xiaomi-run comparisons, not an outside leaderboard. Xiaomi characterizes the release as an open audio-language system that combines general audio understanding with spatial awareness from end to end.
Its own synthetic tests point to the difficult boundary. Xiaomi reported 80.6% overall spatial-question-answering accuracy with one overlapping source, falling to 66.9% with two. Spatial counting fell from 65.2% to 39.5% across those conditions. That weakness is more useful to a prospective developer than an abstract claim of spatial awareness, because multiple simultaneous sounds are precisely where a room or cabin becomes hard to interpret.
The release shares a theme with Tencent’s OmniCapBench tools for testing audio-visual captions: open tooling can make model claims easier to scrutinize. MiDashengLM-Spatial is a preprint available on arXiv, not peer-reviewed validation, and no independent hands-on evaluation was found in the materials reviewed. Its next meaningful test is therefore clear. An outside lab or developer should reproduce Xiaomi’s STAR-Bench and MMAU-Pro spatial subsets on held-out real binaural recordings, swap left and right channels, and report failures with two or three overlapping sources. Adding noisy smart-home and vehicle-cabin recordings would test the proposed use cases where they matter, while also revealing latency, hardware needs, microphone geometry, and behavior with imperfect input, none of which Xiaomi’s examined materials quantify.