AI models

Liquid AI LFM3: six small open AI models for text, images, video and voice on your own device

On October 27, 2026 Liquid AI released LFM3, a family of six open-weight AI models of about 2 billion parameters, all built from the same base. Liquid AI LFM3 reads text, images, video and audio, can also reply in speech, and the text model takes about 1GB: it is designed to run on phones, laptops and small devices without the cloud. Here is what each model does, what the license says and how to try it.

Published on 7 min readBy the editorial team of The AI Creator System

Leggi questo articolo in italiano

AI-generated cover: a small glowing cyan and violet chip floating above a smartphone, with waves of text, images and voice and the title Liquid AI LFM3
Cover image generated with artificial intelligence. How we use AI

Liquid AI LFM3: what was released and when

Liquid AI LFM3 is the third generation of Liquid AI's Liquid Foundation Models, released on October 27, 2026. The official announcement is a post by the Liquid AI team on the Hugging Face blog; the next day, October 28, 2026, TechTimes covered the launch.

What sets LFM3 apart is that six different models start from the same base, LFM3-1.7B-Base, which Liquid AI says was trained on 22 trillion tokens. From that base the company derived specialized versions for text, images and video, voice, all formats at once, and structured decisions. They are all small, between 1.7 and about 2 billion parameters, compared with the hundreds of billions in frontier models such as Gemini 4 Carbon.

The goal is not to beat the giant models but to run AI directly on the device: a phone, a laptop, a car, a robot. Your data stays local, answers arrive without waiting for a server and you do not pay per request.

The six LFM3 models: what each one does

Each LFM3 model has a specific job. Here is the lineup as listed on Liquid AI's blog:

ModelInputOutput
LFM3-1.7Btexttext
LFM3-VL-2Btext, images, videotext
LFM3-Audio-2Btext, audiotext, speech
LFM3-Omni-2Btext, images, video, audiotext, speech
d1-2Btext, images, videoanswers with calibrated probabilities
d1-omni-2Btext, images, video, audioanswers with calibrated probabilities

A few more details:

  • LFM3-1.7B is the text model, with a context of up to 256,000 tokens and about 1GB on disk at 4-bit. A base version, LFM3-1.7B-Base, is also available for anyone who wants to train it on their own data.
  • LFM3-VL-2B has 1.9 billion parameters and is the first vision model in the series that also understands video.
  • LFM3-Audio-2B handles speech recognition and synthesis, voice chat and translation across 18 languages.
  • LFM3-Omni-2B is Liquid AI's first omni model: it accepts every format and replies in text or speech.
  • d1-2B and d1-omni-2B do not write free-form text: they return structured answers with an estimate of how confident they are, which is useful for systems that make automatic decisions.

How to try LFM3 on your computer or phone

The LFM3 weights are a free download from Liquid AI's LFM3 collection on Hugging Face. The LFM3-1.7B model card lists the engines it runs on:

  • GGUF with llama.cpp, Ollama and LM Studio, the easiest way to try it on a PC or Mac;
  • 4-bit MLX for Macs with Apple Silicon;
  • ONNX Runtime for building it into apps;
  • vLLM, SGLang and Transformers for Python and server work.

At about 1GB in 4-bit, the text model fits comfortably on a laptop without a dedicated graphics card. Liquid AI's research mentions running on phones and on a Raspberry Pi 5, but it does not publish minimum RAM requirements, so treat these figures as a guide and test on your own device.

The text model card lists 16 languages: English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian and Polish. For customization, Liquid AI also released Colab notebooks for training (SFT, DPO, GRPO) and says fine-tuning is possible on a single GPU.

LFM3 license: can you use it for work?

Yes, if your business makes less than 10 million dollars a year. The LFM3 models use the LFM Open License v1.0, which allows commercial use only for entities below an annual revenue threshold of 10 million US dollars. Above that threshold, commercial use is not covered by the license and requires a separate agreement with Liquid AI. The threshold does not apply to qualified nonprofits using the models for non-commercial or research purposes.

Other terms worth knowing:

  • if you redistribute the model, you must include the license and state the changes you made;
  • you may not use the Liquid AI trademark except to describe where the model comes from;
  • the model is provided "as is", without warranties.

It is a more restrictive license than OpenMDW 1.1, which Nvidia uses for Nemotron 3.5 and which has no revenue threshold. For a creator or a small business, though, the 10 million cap changes nothing in practice. The license text does not explain how revenue is calculated, so if you are close to the threshold, ask a lawyer.

LFM3 benchmarks: what Liquid AI claims

Every benchmark result published so far comes from Liquid AI and has not yet been reproduced by independent researchers, as TechTimes also points out. These are the main figures claimed in the official blog:

  • LFM3-1.7B scores 38.42 on WebArena-Verified, a test of tasks carried out in a browser, versus 3.02 for the previous LFM2.5-2.6B;
  • LFM3-VL-2B averages 71.4 across 62 vision benchmarks, versus 57.6 for the previous LFM2.5-VL-3B, which has more parameters;
  • LFM3-Audio-2B reaches 3.83 on the UTMOSv2 speech-quality scale, versus 3.22 for the previous audio model;
  • LFM3-Omni-2B is said to beat models of up to 10.7 billion parameters, such as Qwen2.5-Omni-7B, on three audio-video understanding tests, but the blog gives no exact scores;
  • d1-2B scores 75.5 on ImajevBench versus 64.0 for the previous d1-3B, with a third fewer parameters.

The biggest jump is on WebArena-Verified: a 1.7-billion-parameter model completing browser tasks is new, but 38 out of 100 also means it fails more often than it succeeds. Read these numbers as a trend, not as a guarantee for your own use case.

What it means for you: creators and small businesses

For creators and small business owners, LFM3 makes one thing clear: an AI assistant that understands text, images and voice is starting to fit in a file of a couple of gigabytes. It does not replace Claude, ChatGPT or Gemini for complex work, but it opens up new uses:

  1. Offline drafts and summaries. On the road or with confidential data, a 1GB model in LM Studio or Ollama can summarize notes, rewrite messages and organize ideas with no connection.
  2. First-line customer support. A small model can sort frequent questions or draft replies for a person to check: with clear rules, not on its own.
  3. Local voice. LFM3-Audio-2B and LFM3-Omni-2B reply in speech, which is interesting for voice assistant prototypes, although professional-quality voices still call for dedicated tools such as Eleven Realtime.
  4. Check the license. Below 10 million dollars in revenue you can use it for work; if you resell an app built on LFM3, read the redistribution terms.

For step two you need well-written instructions: the 60 prompts in AI Customer Care cover replies on email, chat and WhatsApp, reviews, FAQs and the rules for setting up a chatbot. To test a local model on real office tasks, the 80 prompts in Claude Business OS make a good benchmark, even though they are written for Claude. And if you want to find your way among the tools for creating and selling digital products, there is the AI Business Toolkit.

To choose creative AI tools, start with our AI guides, for example the one on Gemini Omni. For a complete method there is The AI Creator System, and here is how we use artificial intelligence.

Frequently asked questions

When was Liquid AI LFM3 released?

Liquid AI released LFM3 on October 27, 2026 with a post on its Hugging Face blog, and the weights were available right away. TechTimes covered it on October 28, 2026.

Is LFM3 free?

Yes, the weights are a free download from Hugging Face. Run it locally and you only pay for your device; commercial use is allowed for businesses below 10 million dollars in annual revenue.

Which languages does LFM3 support?

The official LFM3-1.7B model card on Hugging Face lists 16 languages, including English, Italian, French, German, Spanish, Japanese, Chinese and Hindi. The audio model translates across 18 languages.

What hardware do I need to run LFM3?

Very little: the LFM3-1.7B text model takes about 1GB at 4-bit and Liquid AI targets phones, laptops and small devices. You can try it with LM Studio, Ollama or llama.cpp; Liquid AI does not publish minimum RAM requirements.

Can LFM3 replace ChatGPT or Claude?

Not for complex work: it is a model of about 2 billion parameters, built for quick, local tasks. It is useful for drafts, summaries, sorting questions and voice prototypes without the cloud.

Read next

Sources checked on October 29, 2026