Liquid AI adds an omni-modal model to the LFM3 family

Background: one model, several senses

Most deployed AI models are specialists. A text model reads and writes text, a speech model turns audio into text or text into audio, and a vision model works on images. An omni-modal model handles several of these modalities inside one set of weights, so a single model can take a picture and a spoken question and answer out loud. "Modality" here simply means the form the data takes: words, pixels, sound.

The LFM3 family from Liquid AI already spans several kinds of model — base, text, vision, audio and decision models ^[1]. Each of those was built around a particular kind of input or a particular job.

What was released

On 21 October 2026, Liquid AI published LFM3-Omni-2B, the first omni-modal entry in the LFM3 family ^[1]. The name follows the family convention, and the "2B" points at its size class: the model has 2.17B parameters ^[1]. Parameters are the learned numbers inside a network; more of them generally means more capacity and more memory, which is why the count is usually quoted alongside a model's name.

Inputs and outputs

The model accepts text, images, video and 16 kHz audio as input ^[1]. The 16 kHz figure is a sampling rate: the audio is described 16,000 times per second, a common setting for speech.

On the output side it returns text, 24 kHz speech, or interleaved text and speech ^[1]. The output speech is sampled at a higher rate than the input audio. "Interleaved" means the reply is not forced into one channel: written text and spoken segments can alternate within a single response, so a model can say part of an answer out loud and write another part.

Size and where it runs

At 2.17B parameters, the model is aimed at phones, laptops and other edge hardware ^[1]. Edge hardware means the device in front of the user rather than a cluster in a data centre. Running a model locally changes what is practical: processing can happen without a network round trip, and the data being processed does not have to leave the device. The trade-off is that the model has to fit within the memory and power budget of a phone or laptop, which is why parameter count matters so much for this class of release.

Quantised and audio-only variants

On 27 October 2026, quantised MLX, GGUF and audio-only variants were uploaded ^[2]. Quantisation stores a model's weights in lower-precision numbers, which shrinks the memory footprint and usually speeds up inference at some cost in accuracy. The two formats named point at different local runtimes: MLX is Apple's array framework for its own silicon, while GGUF is a file format used by a widely adopted family of local inference tools. Uploading both means the same model can be run with the tooling people already have on their machines, on Apple hardware and elsewhere. The release also included variants labelled audio-only ^[2].

Why it matters

Putting several modalities into a 2.17B-parameter model, and shipping formats that run on consumer hardware, makes a different set of applications plausible: a laptop program that watches a video and narrates it, or a phone assistant that listens at 16 kHz and answers in speech at 24 kHz without contacting a server ^[1]. Within the LFM3 family, the release adds a model that crosses the boundaries the earlier text, vision and audio models each kept to ^[1]. The staggered arrival of the quantised and audio-only builds a week after the initial publication also shows how a release now extends past the model weights themselves to the formats and variants that decide where it can actually run ^[2].