Liquid AI's LFM3 Puts Six Specialist Models on One Small Backbone

Key points

• Liquid AI's LFM3-1.7B outperforms its predecessor LFM2.5-2.6B using 35% fewer parameters, according to the company's own benchmarks.
• Six models (text, vision with video input, audio, omni and two decision-making variants) all branch from one shared backbone in the 2-billion-parameter range.
• The text model fits in roughly 1 GB of memory at 4-bit precision, meaning it can run on a phone or a Raspberry Pi.
• Every model in the family can be fine-tuned on your own data using a single consumer GPU.
• Liquid AI published all six models on Hugging Face alongside full model cards and ready-to-run code.

Liquid AI quietly posted the LFM3 family to Hugging Face, the open platform where developers share and test AI models, and it is one of the more interesting small-model releases of the year. Six models, one shared backbone, and a stated goal of running useful AI entirely on the device in your pocket or on the factory floor, no cloud required.

The headline number is the efficiency jump. LFM3-1.7B, the text-only version, beats the previous LFM2.5-2.6B on most tasks while carrying 35% fewer parameters. Parameters are roughly the number of internal settings a model learns during training; fewer usually means smaller and faster. The biggest leap shows on agentic tasks, where an AI agent (software that carries out multi-step tasks on its own) has to click through web pages or operate business tools. On a standard browser-task benchmark called WebArena-Verified, LFM3-1.7B scores 38.42 against its predecessor's 3.02. That is not a typo.

What does "six models from one backbone" actually mean for you?

Instead of downloading one giant general-purpose model, developers pick the version that matches their inputs. Need voice handling? LFM3-Audio-2B covers speech recognition, synthesis and translation across 18 languages. Need to understand video frames for a robotics project? LFM3-VL-2B is the first LFM vision model to accept video. Want the full set? LFM3-Omni-2B takes text, images, video and audio all at once.

Two models in the family, d1-2B and d1-omni-2B, are built for calibrated answers, meaning they are tuned to say "I don't know" rather than guess. Useful for medical or legal tools where a confident wrong answer is worse than a shrug.

For a concrete example: a GP's receptionist app could run LFM3-Audio-2B on a laptop, transcribe a patient's spoken symptoms, and pass structured notes to a booking system, all locally, without sending voice data to a third-party server.

Should privacy matter to ordinary users here?

Yes, and it is one of the more honest selling points. Because these models are designed to run on-device, your data does not have to leave your hardware. That is different from most consumer AI tools, which process your input on remote servers. Liquid AI has not published a formal privacy policy for LFM3 specifically, so developers building products on top of it still need to think carefully about what they log and store. The raw capability, though, makes local processing realistic in a way it wasn't 18 months ago.

All six models run through the standard transformers library, the most widely used toolkit for loading AI models in Python. Fine-tuning, meaning retraining on your own documents or terminology, requires only a single GPU rather than the racks of specialised chips that training from scratch demands.

The d1-2B and d1-omni-2B models in this release build on the decision-model work we covered on 7 October, when Liquid AI's earlier d1 models answered in under 50 milliseconds on a $250 board. LFM3 brings those ideas onto a shared backbone that covers every modality at once, which is the real step change here. The benchmark that matters isn't the efficiency score; it's whether the agentic performance holds up in real deployments. Watch that.

Common questions

Do I need expensive hardware to try LFM3?

About 1 GB of memory at the most compressed setting is all the text model needs, putting it within reach of most modern phones and single-board computers. The omni model that handles video and audio together will need more, but Liquid AI says all six variants target consumer-grade edge hardware.

Is LFM3 free to use?

Liquid AI published the models on Hugging Face with open model cards and runnable code. Commercial licensing terms vary by model; check the individual model cards on Hugging Face for the exact licence before building a product on top of one.

What is 4-bit precision and why does it matter?

4-bit precision is a compression technique that reduces a model's stored numbers from 32 or 16 bits down to 4, shrinking the file and the memory needed to run it at a small cost to accuracy. It is the main reason LFM3-1.7B fits in roughly 1 GB.