Models
What Is LFM3? Liquid AI’s Tiny Open Models for Phones, Explained
LFM3 is Liquid AI’s new family of six small open-weight models, released October 27, 2026. What each one does, the 1GB text model, the license and how to try it.
LFM3 is the newest family of AI models from Liquid AI, and it went from zero to the top of Hacker News in a day. If you saw "LFM3" or "Liquid AI LFM 3" in your feed and wondered what it is, the short version is: small AI models built to run on a phone or laptop instead of a data center.
Below is what LFM3 is, the six models in the family, what the 1GB claim really means, whether you can use it for free, and how it stacks up against bigger names like Qwen and Gemma.
What is LFM3?
LFM stands for Liquid Foundation Model. Liquid AI started the line in 2024 and shipped LFM2 and LFM2.5 since then. LFM3 is the third generation. Liquid AI calls it "every modality for specialized intelligence," which means one shared base model that branches into versions for text, pictures, sound and decisions.
All six models grow from the same root, called LFM3-1.7B-Base. It was trained on 22 trillion tokens of data. Liquid AI tested the design on 14 combinations of real devices and software, including the Samsung Galaxy S26 Ultra and OnePlus 15 phones, a Raspberry Pi 5, an Apple M4 Max laptop and AMD Ryzen AI chips. The goal was speed on normal hardware, not a record on a giant server.
The six LFM3 models
| Model | What goes in | What comes out | Size |
|---|---|---|---|
| LFM3-1.7B | Text | Text | 1.73B, 256K context, 16 languages |
| LFM3-VL-2B | Text, images, video | Text | 1.9B |
| LFM3-Audio-2B | Text or audio | Text or speech | About 2B, 18 languages |
| LFM3-Omni-2B | Text, images, video, audio | Text and speech | 2.2B |
| d1-2B | Text and images | A calibrated answer or score | About 2B |
| d1-omni-2B | Text, images, video, audio | A calibrated answer or score | About 2B |
Two firsts stand out. LFM3-VL-2B is the first Liquid vision model that can watch video, and LFM3-Omni-2B is the company’s first model that takes in every type of input at once. The two d1 models do not chat at all. Like Microsoft Decision-1, they pick an answer and say how sure they are.
Does LFM3 really run on a phone?
The text model does, on paper. At 4-bit precision (a compressed format), LFM3-1.7B takes about 1GB of memory. Liquid AI also says a full 128K-token conversation needs only 1.6GB of extra memory for its working cache, against 4.3GB for Qwen3.5-4B in the same test. That is why long chats do not crash small devices as quickly.
On three of its test devices, Liquid AI says LFM3-1.7B writes text 1.8 to 2.8 times faster than the next-fastest model it compared. These are the company’s own numbers.
LFM3 benchmarks: how good is it?
Liquid AI published scores against bigger open models. All of these are self-reported on launch day, so treat them as a first look, not a final verdict.
| Test | LFM3-1.7B | Gemma 4 E4B | Qwen3.5 4B | Qwen3.5 9B |
|---|---|---|---|---|
| BFCL V4 (tool calling) | 71.39 | 48.25 | 55.12 | 62.83 |
| WebArena-Verified (web tasks) | 38.42 | 17.06 | 21.92 | 28.94 |
| IFEval (following instructions) | 96.66 | 90.02 | 92.52 | 92.08 |
| GPQA Diamond (hard science) | 63.37 | 57.88 | 75.76 | 80.40 |
| SWE-bench Verified (coding) | 43.40 | 0.50 | 41.20 | 51.20 |
| Average of 29 tests | 49.37 | 31.74 | 39.27 | 47.24 |
The pattern is clear. LFM3 is strong at agent-style jobs: calling tools, following instructions and clicking through websites. It is weaker at deep knowledge and coding, where models four to five times its size still win. Liquid AI says so itself: the model card says LFM3-1.7B is not recommended for agentic coding or complex multi-step reasoning because of its small size.
Is LFM3 free?
Mostly. The weights are free to download from Hugging Face under the LFM Open License v1.0. According to an analysis by i-scoop, the license is based on Apache 2.0 but adds a revenue line: commercial use is free for organizations with less than $10 million in annual revenue, research and non-commercial use have no limit, and bigger companies need a separate deal with Liquid AI. Liquid AI has not published a price for that deal.
How to try LFM3
- In your browser: the Liquid AI Playground at playground.liquid.ai.
- On a computer: download a GGUF file for llama.cpp, Ollama or LM Studio, or the MLX 4-bit version for Apple Silicon Macs.
- For developers: the Liquid AI API, OpenRouter (a launch partner), or vLLM and SGLang on a GPU. Every model can be fine-tuned on a single GPU.
There is no official consumer LFM3 app for iPhone yet. Running it on a phone today means a developer tool or a third-party app that loads GGUF files.
LFM3 vs other small models
LFM3 joins a crowded field of "small but useful" models. Nvidia’s Nemotron 3.5 Bolt E4B targets about 4GB on Nvidia graphics cards for voice assistants and robots. LFM3-1.7B aims lower, at about 1GB, and covers more device types, from phone chips to Raspberry Pi. Google’s Gemma 4 E4B and Alibaba’s Qwen3.5 4B are the closest rivals, and on Liquid AI’s own chart LFM3 beats both on average while being smaller.
The launch partner list is long for a small lab: AMD, Arm, Intel, Nvidia, Qualcomm, OpenRouter, Vercel and Ambarella supported it on day one, and Liquid AI says Mercedes-Benz and Shopify helped test it before release.
What this means for you
LFM3 is a sign of where AI is going: smart enough to call tools and fill in web forms, small enough to live on your phone without the cloud. That will mean faster, more private assistants in cars, shopping apps and gadgets. But today it is a building block for developers. A 2-billion-parameter model will not plan a trip or research a decision as well as the big cloud models, and setting it up takes work.
If you want an AI that finishes real tasks right now, an agent app gets you there without picking or installing models. Agentic AI gives you a team of expert agents on your iPhone and iPad that split a job into steps and work through it together. Not sure what an AI agent is? Start with what agentic AI means.
New to terms like open-weight or tool calling? See our agentic AI glossary.
Quick answers
What is LFM3? A family of six small open-weight AI models from Liquid AI, released October 27, 2026, covering text, vision, audio, omni and decision tasks. Each is about 2 billion parameters.
Who makes LFM3? Liquid AI, a startup spun out of MIT. LFM3 is the third generation of its Liquid Foundation Models.
How much memory does LFM3 need? The text model, LFM3-1.7B, needs about 1GB at 4-bit precision, according to Liquid AI’s model card. That is small enough for most recent phones and laptops.
Is Liquid AI LFM3 open source? It is open-weight under the LFM Open License v1.0. Commercial use is free below $10 million in annual revenue; above that you need a license from Liquid AI.
Can I run LFM3 in Ollama or LM Studio? Yes. Liquid AI ships GGUF files for llama.cpp, Ollama and LM Studio, plus MLX for Macs and ONNX.
Is LFM3 better than Qwen3.5? On Liquid AI’s own 29-test average, LFM3-1.7B scored 49.37 against 47.24 for Qwen3.5 9B. Qwen still wins on hard science and coding tests.
Sources: Liquid AI: Introducing LFM3 · LFM3-1.7B model card · Unite.AI · MarkTechPost · i-scoop on the license
