Liquid AI

What Liquid AI is

Liquid AI describes itself as an efficiency-first foundation model company ^[1]. The point of that label is that the cost of running a model is treated as a design constraint rather than something fixed afterwards: the company's public argument is less about serving models from a cloud API and more about models that run on hardware people already own ^[1].

Where it came from

The company says it was spun out of MIT CSAIL by four founders: Ramin Hasani as chief executive, Mathias Lechner as chief technology officer, Alexander Amini as chief scientific officer, and Daniela Rus ^[1]. It traces its technical lineage to liquid neural networks, STAR and the LFM family, which it describes as modern recurrent systems that can improve their own performance ^[1]. A recurrent system is one that reads an input step by step while carrying an internal state forward, rather than treating each input as unrelated to the last.

The argument

Its stated thesis is that intelligence has to be ubiquitous, and that cloud-only serving is intractable on cost, power and latency ^[1]. The scale it points to is device shipments: nearly 2.2 billion devices that can host a foundation model ship every year across phones, PCs, wearables and cars, before robots and machines are counted ^[1]. Because memory, battery and latency limits differ across CPUs, GPUs and NPUs, the company argues that foundation models must be co-designed with hardware ^[1].

The LFM3 family

On 21 October 2026 Liquid AI published LFM3-Omni-2B, its first omni-modal model in the LFM3 family ^[2]. Omni-modal means one model accepts several kinds of input rather than one: this model takes text, images, video and 16 kHz audio, and returns either text or 24 kHz speech ^[2]. It has 2.17 billion parameters, supports 23 languages, and carries the custom LFM 1.0 licence rather than an OSI licence ^[2]. Quantised variants — compressed builds that trade numerical precision for smaller size and lower memory use — were uploaded for the MLX and GGUF runtimes on 27 October 2026, together with audio-only variants ^[3].

Where it sits

Two details of that release say something about the company's position. The first is size: at 2.17 billion parameters the model is small enough to be a candidate for local deployment, which is the direction the company keeps pointing at ^[1][2]. The second is licensing: LFM 1.0 is a custom licence rather than an OSI licence, so the terms on which others may reuse the model are set by the company rather than by a standard open-source licence ^[2]. The release pattern follows the same logic. The base weights appeared first, and within a week came quantised builds aimed at common local runtimes and an audio-only version ^[2][3]. Taken together, the company's claim is that a foundation model can be small, multimodal in both directions, and distributed as downloadable weights — with the licence, not the architecture, deciding who gets to build on it.