Liquid AI's phone-sized model promises deployment "without restrictions" — its licence sets one at 10 million dollars

Liquid AI published LFM2.5-VL-3B on 12 August 2026, and the engineering claim behind it is genuine. The model has 3.12 billion parameters, occupies about 3 GB of memory and reads screens, documents and charts on hardware the reader already owns: the company measures 228 tokens per second on an Apple M5 Max, 116 on an AMD Ryzen AI Max+ 395 and 20 on a Galaxy S26 Ultra — the last of those a phone with no cloud behind it. Across 28 vision benchmarks Liquid AI reports an average of 69.4, level with InternVL-3.5-4B and marginally behind Qwen3.5-4B, both of which are larger models. Those figures come from the vendor and have not been independently reproduced. The company's announcement invites readers to "download, fine-tune, and deploy without restrictions". The licence shipped with the weights says something narrower. LFM Open License v1.0 is not an open-source licence in the usual sense: alongside familiar Apache-style redistribution terms it adds a section headed "Commercial Use Limitation", under which the commercial grant applies only to entities whose annual revenue stays below 10 million US dollars. Above that threshold, in the licence's own words, commercial use "is not licensed under this Agreement" — a separate arrangement with Liquid AI is required. Registered non-profits are exempt, but only for research and other non-commercial purposes. The practical effect depends entirely on who is reading. For a student, a hobbyist or a small studio the model is free in every sense that matters, and running it on a phone rather than renting a data centre is precisely the point. For a mid-sized company the threshold is close enough to matter, and for a large one the published licence does not cover production use at all. The contrast with the past fortnight is instructive. On 14 August Alibaba published Qwen3.8-27B under Apache 2.0, with no revenue condition attached, while keeping its largest model closed and cloud-only — a split that is at least legible from the outside. Liquid AI has taken the other route: one licence for everyone, with the limit written into the text rather than into what is withheld. Both are now routinely described in coverage as "open weights", a phrase that has quietly come to cover terms that differ sharply in what they permit. Nothing here diminishes the model. It is a well-made piece of engineering aimed squarely at devices, and we have given it a profile on that basis. It is simply worth knowing which of the two sentences — the announcement's or the licence's — governs a given reader's use.