You give it a photo, a video clip, a recording or plain text, and ask about it by typing or by speaking. It answers in text, in speech, or both at once. Hold up a photo of a street sign and ask what it says, play it a clip and ask what the person in it just said, or simply talk to it like an assistant and get a spoken reply.
The difference is that one model handles the whole loop. With a video that has sound, it watches and listens at the same time, so it can answer questions that need both, such as who is speaking when the door slams. The same model also transcribes speech, reads text aloud in a male or female voice, and translates speech into text in another language.
It is small: about 2 billion parameters, built to run on laptops and phones. In Liquid AI's own tests it beat larger models from Qwen, Google and Microsoft at answering questions about video with sound. It understands text in 16 languages and speaks 19, Turkish among them.