Domux: a compact open model for smart-home command understanding at the edge

Voice and chat assistants for the home face the challenging task of transforming ambiguous natural language into clear, organized commands. For instance, the phrase "Make it cozy in here" must be converted into a specific intent and relevant slots, such as identifying the device, room, and value involved. Domux is an open-source model developed by iFlytek that specifically addresses this issue, focusing on command understanding for smart-home assistants through intent parsing and slot filling.

Domux's primary task is to comprehend smart-home commands, breaking them down into intent parsing and slot filling. Its base model is fine-tuned on google/gemma-4-E2B-it, designed for multimodal input, combining both images and text. The model is intended for deployment at the edge, on devices rather than relying on large cloud models.

This is crucial, as keeping command understanding on the device can reduce communication round-trips and keep more interaction local, instead of transmitting every utterance to a massive centralized model. The compact nature of the base model used in Domux is important because it allows the model to run in a size class suitable for deployment close to the device.

This is particularly appealing for home assistants, as it enables more efficient and localized command understanding. To try Domux, you can find the model card on Hugging Face, though access is restricted and you may need to log in and request permission. The link is: https://huggingface.co/iFlytekOpenSource/Domux. The sharing of open work like Domux is encouraged because on-device, task-specific models are an essential part of the foundation model and serving story. It's not necessary for every task to involve a massive cloud-based model.