Liquid AI released LFM3 on October 27, 2026, giving software teams a downloadable model for local assistants that handle spoken questions about pictures or video with sound. It expands the company’s offering beyond separate text, vision and audio models.
LFM3-Omni-2B accepts text, images, video and audio and can answer in text or speech, according to Liquid’s documentation. Its weights—the files needed to run it—are public on Hugging Face. This is a component for developers, not a ready-made consumer assistant.
When sight and sound belong together
Liquid’s examples pair a spoken question with an image, or ask about a clip using both its frames and soundtrack. For a product team, the useful distinction is whether the answer needs sight and sound together. Liquid recommends the separate Audio model for speech-only tasks.
The variants also have different context windows: the budgets for input and output in one exchange, measured in units called tokens. Omni and LFM3-Audio list 32K tokens; the image-and-video model has 128K and produces text; the text-only model has 256K. A larger budget is not permanent memory or a guarantee of how many minutes of video will work.
The software determines what works locally
The model’s full capabilities are not available through every software package. Liquid documents audiovisual input and speech output through Transformers, a library for running models. Its SGLang setup returns text only.
Support for Omni was merged into llama.cpp on October 27, but that change leaves video input for later. Downloadable weights therefore do not mean every local AI application can already handle a video and its soundtrack.
Even in the documented audiovisual workflow, video is sampled at two frames per second and developers must explicitly include the soundtrack. Supplying a video file alone does not necessarily give the model its sound.
Language coverage needs a separate check: Hindi appears in Omni’s text-language list but not its published speech list. Text support should not be treated as a promise of voice support.
Liquid’s roughly 1 GB memory claim concerns the text model at reduced numerical precision, not a complete Omni application. As with Liquid’s earlier local assistants, teams still need to establish the memory requirement and response speed on their intended device.
Commercial use has conditions
Liquid’s licence guide says commercial use is free for organisations with annual revenue under US$10 million. At US$10 million or more, a separate commercial licence is required. The relevant legal entity includes businesses under common control, not just the project using the model.
The guide also says downstream customers using the models commercially must assess their own eligibility. A business still has to budget for hardware, integration and ongoing maintenance even when the model licence is free.