Liquid AI's New Models Put a Capable Assistant on Your Phone, No Cloud Required
Liquid AI's LFM3 family brings text, vision and voice AI onto phones. Your data stays on your device, but big companies will need a paid license.
AI that answers you without sending your words to a data center just got more practical. On 27 October, Liquid AI released LFM3, six open models of about 2 billion parameters each. That is small enough to run on a phone, a laptop or a Raspberry Pi. When the model runs on your device, your questions, photos and voice stay there too.
What Liquid says it can do
The family handles text, images and audio, and two models are built for making decisions. One "omni" model can see, listen and talk back in a single model. Liquid says it tested the models on 14 device and hardware setups. These include the Samsung Galaxy S26 Ultra, the OnePlus 15, Apple M4 Max and AMD Ryzen AI laptops, and Qualcomm's phone AI chips.
Liquid's speed figures are its own and haven't been checked by anyone else. The company says its 1.7 billion parameter text model writes about 43.6 tokens a second on an iPhone 17 Pro, using under 2GB of memory. That is faster than most people read. On its own blog, Liquid says the text model ran 1.8 to 2.8 times faster than the next fastest rival on each of the devices it tested, and that compressed it fits in about 1GB. Treat these as company claims until someone tests them independently.
Liquid also says the text model can handle very long inputs, up to 256,000 tokens, and is trained to work as an agent: sorting email, managing a calendar or researching on the web. The audio model supports 18 languages, according to the company.
Built to be customized
Liquid's pitch is that small models work best as specialists. It says every LFM3 model can be retrained on a company's own data using a single graphics chip. In one demo, Liquid says a few hours of training turned its vision model into a tool that spots defects in drone photos of wind turbine blades. Mercedes-Benz says it is exploring the models for its in-car voice assistant, with fast, private, cloud-independent features as the goal. That is a partner statement, not a product announcement.
The fine print for business
The models are free to download, modify and keep private. But Liquid's license isn't fully open source. Commercial use is free only for companies with less than $10 million in yearly revenue. That total includes parent companies and affiliates. Bigger firms need a paid license from Liquid. Research and qualified nonprofits are exempt.
For startups and developers, that's a low-cost way to build apps that keep user data on the device. Larger enterprises get a privacy-friendly option too, but they will have to pay for it.
Small models still have limits. They are weaker than big cloud models at long reasoning and complex tasks. Liquid's own tables show larger models still win on coding and general knowledge.
What to watch: independent speed and accuracy tests, and whether phone makers or app developers ship LFM3 inside real products.
Sources
- Introducing LFM3: Every Modality for Specialized Intelligence (Liquid AI)
- Build specialized multimodal models with LFM3 (Hugging Face)
- LFM License (Liquid AI)
- LFM Open License User Guide (Liquid AI Docs)
- LFM3: a 2B model that runs on your phone (CLD Navi)
- LFM3-1.7B (Qualcomm AI Hub)
Photo: "Hand holding smartphone with blank white screen" by Santeri Viinamäki, via Wikimedia Commons, CC BY-SA 4.0. Graded and overlaid by PSI, shared under the same license.