Liquid AI's LFM3 Puts Six Specialist Models on Your Phone

Liquid AI released a family of six open-weight models on October 27, each built around two billion parameters and small enough to run on a phone. Called LFM3, the set splits the work by job rather than scaling one model up: separate releases handle text, vision, audio, an omni mode that takes several inputs at once, and decision tasks for agents.

The appeal is the footprint. According to unite.ai, the 1.7-billion-parameter text model fits in roughly a gigabyte at four-bit precision and carries a 256,000-token context window, enough to keep a long document in view on a device with no connection to a data centre.

A different shape of model

Under the hood, LFM3 does not look like a standard transformer. MarkTechPost describes a tri-hybrid backbone that mixes three kinds of operators across its layers, leaning on gated short convolutions and sliding-window attention with only a handful of global attention layers. The design trades some of the brute-force reach of full attention for speed and a smaller memory bill, which is the whole point when the target is a handset rather than a server rack.

Liquid says its 1.7B text model averages 49.37 across 29 benchmarks, edging out a larger Qwen model on the same tests. Those figures are self-reported, as benchmark numbers from a model's own maker usually are, so they are worth checking against independent runs before taking them as settled.

The on-device argument

The models are available on Hugging Face, through Liquid's playground and its API, with the text model shipping under an open licence. The strategy sits at the opposite end of the industry from the trillion-parameter frontier. Rather than renting the biggest possible model over a network, LFM3 bets that a specialist model running locally is often enough, and that privacy, latency and cost all improve when the work never leaves the device.

It is a crowded bet. The push to shrink capable models onto everyday hardware has become one of the busier corners of open-weight research, and it feeds directly into the question of what you can realistically run yourself. A two-billion-parameter model on a phone is not a frontier system, and Liquid is not claiming otherwise. It is claiming that for a lot of ordinary tasks, you may not need one.