Liquid AI has released six open LFM3 models spanning text, vision, audio and decision-making, with some designed to run locally on phones and laptops. The standout practical detail: GIGAZINE says the LFM3-1.7B ran at 43.6 tokens per second on an iPhone 17 Pro while using 1.90GB of memory.
Watch Desk analysis
What happened
The initial release includes LFM3-1.7B for text; LFM3-VL-2B for text, image and video input; LFM3-Audio-2B for text and audio; and LFM3-Omni-2B, which handles text, images, video and audio. Two more, d1-2B and d1-Omni-2B, are decision-making models that can process visual inputs, with d1-Omni-2B also handling audio.
GIGAZINE reports that all six are open models. It also says LFM3-VL-2B outscored Qwen3.5 2B and Gemma 4 E2B in cited benchmarks, and surpassed the larger Gemma 4 12B in some tests. Those are reported benchmark results, not independent testing.
Why it matters
The release puts several kinds of AI, not just a text chatbot, into a family aimed at smaller devices. Running a model locally can make on-device use possible without sending each task to a cloud service, while the reported phone speed and memory figure give readers a tangible sense of what one model may require.
It is also a reminder that “small model” does not mean one thing. A text model, an audio model and a decision model have different jobs, and the practical question is whether each does its particular job well enough to be useful.
Our read
The most interesting part is the range: language, audiovisual input and models that output decisions are arriving together in a compact family. The phone result is encouraging, but one device and one model do not settle how well the whole suite runs across real workloads. Still, local AI is becoming less of a laptop-only parlour trick.
What to watch
- Whether Liquid AI publishes comparable performance and memory figures for the other models.
- How the models perform on everyday tasks beyond the cited benchmarks.
- Whether developers build useful on-device applications around the decision models.
Discussion spark: For on-device AI, would you prioritise a model that handles several kinds of input, or one smaller model that does a single job especially well?
Sources and evidence
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.