Is LFM3 Safe? Liquid AI’s Open Models and Your Data

Short answer: it depends almost entirely on where you run it. LFM3 is a family of small open-weight models, so you can download them and run them on your own laptop or phone, and then what you type never reaches Liquid AI at all. Use Liquid’s own Playground or API instead and its privacy policy applies, which allows your inputs and outputs to be used for training. The model itself comes with very little published safety material.

Liquid AI announced LFM3 on 27 October 2026. Its launch post describes six models of around two billion parameters each, all branched from one base model, and says the weights are available on Hugging Face, in the Liquid Playground and through the Liquid API. Unite.AI’s report lists the line-up: a text model, a vision model, an audio model, an omni model that takes all three, and two decision models. This article reads what is documented about data and safety, and what is not.

What LFM3 is, briefly

The text model, LFM3-1.7B, has 1.73 billion parameters and a 256,000-token context window, according to its Hugging Face model card. It supports 16 languages, including German, French, Japanese and Korean. Liquid says it fits in about 1 GB of memory at 4-bit precision, which is why the launch keeps returning to phones, laptops and cars rather than data centres.

That size is the point, and it is also the limit. A model this small knows less than ChatGPT or Claude and gets facts wrong more readily. The model card says so: it is not recommended for agentic coding or complex multi-step reasoning, and it suggests pairing it with retrieval or search for knowledge-heavy tasks. If you are unsure how a chatbot and an agent differ, AI agent vs LLM sorts that out first.

Route one: running it yourself

This is the route that makes LFM3 interesting for privacy. Once the model file is on your machine, it runs there. A model file on its own sends nothing anywhere; the prompt, the document and the answer all stay on the device. For anyone who has hesitated before pasting a client list into a chat window, that is a real difference.

Two cautions. First, the app you run it in is a separate piece of software with its own behaviour. A local chat app, a phone app built on LFM3 or a browser tool may log, sync or send data on its own terms, so check that app, not the model. Second, local does not mean correct. Everything in how to check an AI answer when you are not the expert applies with more force to a two-billion-parameter model than to a large one.

Route two: Liquid’s Playground and API

If you use LFM3 through Liquid’s own services, your text goes to Liquid. Its privacy policy, last updated on 30 September 2026, lists among its purposes training, fine-tuning or otherwise enhancing AI models using the inputs and outputs you share. It says Liquid does not seek to collect personal data for training, but that training data may include it incidentally.

• —Where it is processed: the policy says the services are hosted and operated in the United States, with transfers elsewhere covered by mechanisms such as Standard Contractual Clauses.
• —Your rights in Europe and the UK: the policy lists access, erasure, portability and objection under the GDPR, and says you can object to your personal data being used to train its models.
• —How long it is kept: as long as necessary for the purposes in the policy, varying by service and agreement, with deleted data possibly persisting in backups for a limited time.

None of that is unusual for an AI company, but it is the opposite of the local route. The same rule we gave for uploading Excel to ChatGPT holds here: decide what leaves your machine before it leaves.

What Liquid has not published

The LFM3-1.7B model card we read on 28 October 2026 has no safety, bias or misuse section, and the launch post makes no safety claims. There is no system card of the kind OpenAI publishes for its models, and no outside evaluation yet. That does not mean the models are unsafe. It means the evidence either way has not been published, and anyone telling you otherwise is guessing.

With an open model, the vendor’s privacy policy stops mattering the moment the file is on your device. Its safety testing does not, and for LFM3 there is very little of it on the record.

This is a pattern with open releases. When Meta opened its model, we found a similar split in is Muse Spark Open safe: strong control over your own data, thin documentation of behaviour. Mistral’s documentation for Mistral Large 4 is fuller on data handling, but it is a hosted model, so the question is different.

The licence, if you use it at work

The model card puts LFM3-1.7B under the LFM Open License v1.0. The licence text is based on Apache 2.0 but adds a threshold: commercial use is licensed only for organisations with annual revenue below US$10 million. Above that, you need a separate agreement with Liquid. Qualified non-profits using it for research or non-commercial purposes are exempt from the threshold. Liquid’s licence FAQ explains the details. If your employer is larger than that, ask before you build anything on it.

What to do with it

1. 01If privacy is the reason you are interested, run it locally and check the app you run it in.
2. 02If you use the Playground or API, treat it like any hosted chatbot and keep personal and client data out unless your organisation has an agreement.
3. 03Use it for narrow, checkable jobs — sorting, extracting, drafting short replies — not for facts you cannot verify.
4. 04Write clear, specific instructions. Small models are less forgiving of vague prompts; writing a prompt that works first time covers the habit.

If the audio model is what caught your eye, the questions are the same ones we asked of another open speech model in is Falcon ASR safe: where the recording goes, and what the demo does with it.

Coursium is a mobile app on the App Store that teaches people to use AI at work. New models arrive every week. Knowing where your data goes, writing a clear instruction and checking the answer carry over from one to the next. If that is what you want to practise, have a look at Coursium.