Liquid AI has released LFM3, a new family of six open-weight models built around two billion parameters each, designed to run on devices rather than in the cloud. The family covers text, vision, audio, combined multimodal input and two decision models, all branched from a single shared base model. Liquid says the models are available now on Hugging Face, its own playground and its application programming interface, with weights that developers can download, fine-tune and deploy locally.
One backbone, six specialised models
Instead of one large general-purpose model, Liquid offers six smaller models, each tuned for a different job, all built from the same LFM3-1.7B-Base backbone. That base was pre-trained on 22 trillion tokens, with a final 3.5 trillion-token mid-training stage that added more scientific papers, repository-level code and agentic traces, and extended the context window to 256,000 tokens.
The text model, LFM3-1.7B, is aimed at local agents and specialised applications. Liquid says it can triage an inbox, manage a calendar, research topics on the web and use tool search to find the right tool among thousands of options, and that it can be adapted to a new domain with a short fine-tuning run on a single graphics card.
The claim is not that a small model beats a frontier model at everything, but that it is good enough for bounded, repetitive agent work where data should stay on the device.
Vision, audio and omni
LFM3-VL-2B is the family’s vision model, and the first Liquid vision model to accept video as input. It can answer questions about a recorded lecture, a walkthrough filmed on a phone or a clip from a camera feed, and it reasons step by step before answering harder visual questions.
Liquid reports that across 62 vision benchmarks it scores well ahead of other models of similar size, and close to much larger models on its average score, though those figures are Liquid’s own and have not been independently verified.
It can transcribe, synthesise and translate speech, interpret recordings and turn spoken requests into tool calls that control applications and devices. Liquid says its new audio encoder is fully causal, so it can process live microphone input while a person is still speaking, with a time to the first token as low as 50 milliseconds.
LFM3-Omni-2B, the company’s first omni model, brings vision, audio and language together in a single model that can see, listen and talk back. A user can ask a spoken question about an image, or provide a video with synchronised audio and ask a question that depends on both what is seen and what is heard.
At 2.2 billion parameters, Liquid pitches it for tutors that look at handwritten work, tablets that answer questions about what they see, and in-car assistants that combine camera input with driver speech.
Decision models that answer, not chat
The two decision models are the family’s most unusual members. Called d1-2B and d1-Omni-2B, they do not write an answer at all: given a situation, such as a support ticket, an invoice, an application screen, a photo or a voice note, plus a list of questions, they answer all of them in a single forward pass, each with a calibrated probability.
The newer typed answers can also point to an object, read text and dates from a document, give a number with its likely range, or pick every option that applies.
That trade, smaller and stronger, is the pitch for the whole family: a model that turns a stream of tickets, documents, calls or camera frames into structured decisions without a round trip to the cloud. For teams building agents that must act on private data, that local, structured behaviour may matter more than raw conversational fluency.
Why It Matters
LFM3 lands in a month when frontier models have dominated the news, from Google’s cybersecurity-focused Gemini 4 Carbon to a wave of open-weight releases from other labs. Liquid’s bet is the opposite direction: that the next useful wave of artificial intelligence will be small, specialised and local, running on laptops, phones, edge devices and even a Raspberry Pi.
If that bet is right, the most important models of the next few years may be the ones users never notice, quietly handling tasks in the background.
The open-weight angle strengthens that case. Developers can download the weights, test them on their own data and fine-tune them for a narrow task, as Liquid demonstrated by fine-tuning the vision model to classify defects in drone photographs of wind turbine blades.
That kind of specialisation is hard to buy from a closed frontier model, and it is why open-weight families keep drawing attention even when they cannot match frontier models on broad knowledge. Our report on Kimi’s open-weight launch shows the same pattern from another lab, while Gemini 4 Carbon shows the frontier end of the same month.
Background
Liquid AI is a Massachusetts Institute of Technology spin-off that builds small models designed to run on devices rather than in data centres. LFM3 is its third generation of Liquid Foundation Models, following LFM2.
Source: Liquid AI announcement
0 Comments