Liquid AI launched LFM3: six small models that run on the device, with a license to read before you sell
Direct answer
On October 27, 2026, Liquid AI introduced LFM3: six open-weight models of around 2 billion parameters each, all derived from one shared base, LFM3-1.7B-Base, pretrained on 22 trillion tokens. The bet is direct: a useful agent does not always need a datacenter. It can run on a phone, a laptop or an edge device, and it can be fine-tuned to a specific domain with a single GPU.
The family, specifically
- LFM3-1.7B (text): text and tool calling, with a 256,000-token context and about 1 GB of memory at 4-bit, per the company.
- LFM3-VL-2B (vision): reads images, video, documents and screens. It is Liquid's first vision model to accept video input.
- LFM3-Audio-2B (audio): understands and produces speech across 18 languages, with a reported time to first token as low as 50 milliseconds.
- LFM3-Omni-2B: combines text, images, video and audio in one model.
- Two decision models: aimed at typed decisions on text and images, for agentic tasks rather than conversation.
- Architecture: Liquid calls it tri-hybrid: gated short convolutions, sliding-window attention and global attention, searched against real hardware, including phones and edge devices, not only GPU clusters.
The figures, with their label, and the license
- Company-reported performance: the text model averages 49.37 across 29 benchmarks, above a model more than five times its size, per MarkTechPost's review of the launch. The vision model rises from 57.6 to 71.4 across 62 vision benchmarks against its predecessor. All of those figures are the company's own measurements and lack independent verification so far.
- The limit its own model card states: LFM3-1.7B is not recommended for agentic coding or complex multi-step reasoning, and knowledge-heavy questions should use retrieval or search. A small model wins where the job is acting; it loses where the job is long reasoning.
- The free commercial license has a cap: the LFM Open License v1.0 is based on Apache 2.0 and adds a revenue threshold. Commercial use is free for organizations with less than $10 million in annual revenue. Above that threshold, commercial use is not covered and requires a separate agreement with Liquid. Research and non-commercial use have no threshold.
- The question for your business: if the agent must work without sending data to the cloud, respond fast and fit a specific process, a small local model is a serious option. If the agent must reason through complex cases or write production code, this size, per its own card, is not the right tool.
Want AI that works for your business?
Message us on WhatsApp and let's talk about automation and AI agents for your business.
WhatsApp inicIARelated articles
- The war of cheap verdicts: the AI layer that decides without writing an essay
- What an AI agent is and how a real business uses one
- How to implement AI in your business in 30 days
Frequently asked questions
What did Liquid AI launch on October 27, 2026?
The LFM3 family: six open-weight models of around 2 billion parameters, derived from one shared base, to run and be fine-tuned on local hardware.
Can LFM3 run on a phone?
Per the company, the text model takes about 1 GB at 4-bit, making it viable on phones and laptops without a dedicated GPU.
Are the performance figures verified?
Not independently. They are the company's own measurements from its launch and model cards.
Can I use LFM3 commercially for free?
Only below the license threshold: less than $10 million in annual revenue. Above it, commercial use requires a separate agreement with Liquid.
