If you want a capable AI model that runs on your own phone, laptop, or Raspberry Pi with no cloud account, Liquid AI LFM3 is worth a look. The 1.7B text model needs roughly 1 GB of memory at 4-bit precision, and it handles a 256K-token context window.
Liquid AI released the six-model family in late October 2026. It’s open-weight, so you can download it today. Here’s what it does well, where it falls short, and how to try it.
Here are the facts worth knowing before you download anything.
- Liquid AI LFM3 is a family of six open-weight models: text, vision, audio, omni, and two “decision” variants.
- The text model, LFM3-1.7B, has about 1.73 billion parameters and a 256K-token context window.
- It supports 16 languages and ships under the LFM Open License v1.0.
- At 4-bit precision it fits in roughly 1 GB, according to AI Weekly.
The short version: it’s built for small, fast, private tasks, not heavy reasoning.
Why this small open-weight model matters
Most on-device models force a trade-off. Go small and they stumble on simple instructions. Go bigger and your phone gets hot.
Liquid AI took a different route with the architecture. LFM3 mixes gated short convolutions with sliding-window and global attention layers across 24 layers. In plain terms, most of the model looks only at nearby text, and a few layers look at everything. That keeps memory use low as the context grows.
Liquid AI says the 1.7B model beats its earlier LFM2.5-2.6B while using about 35% fewer parameters. It also claims decoding that is 1.8 to 2.8 times faster than the next model on three on-device setups. Those are the company’s own numbers, so treat them as a starting point and test on your hardware.
What the Liquid AI LFM3 family includes
All six models branch from the same LFM3-1.7B-Base backbone, which was pre-trained on 22 trillion tokens. That shared base means the vision, audio, and omni versions stay in the same small size class.
Liquid AI lists these hardware targets for the family, per AI Weekly:
| Device type | Examples named by Liquid AI |
|---|---|
| Phones | Samsung Galaxy S26 Ultra, OnePlus 15 |
| Single-board computer | Raspberry Pi 5 |
| Laptops and desktops | Apple M4 Max, AMD Ryzen AI |
| NPUs | Qualcomm Hexagon |
Why this matters: a model that runs on a Raspberry Pi will run comfortably on almost any laptop from the last few years.
How to run LFM3-1.7B locally
The official model card lists several ways to run it. Pick the one that matches your setup.
| Your setup | Format and tool |
|---|---|
| Windows, Mac, or Linux with an app | GGUF in LM Studio or Ollama |
| Command line on any PC | GGUF in llama.cpp |
| Apple Silicon Mac | MLX 4-bit |
| Server or GPU box | BF16 in vLLM, SGLang, or Transformers |
| Cross-platform apps | ONNX Runtime |
For most people, the easiest path is a GGUF file in LM Studio or llama.cpp. Search for the LFM3-1.7B GGUF on Hugging Face, download a 4-bit version, and load it.
The model card recommends a low temperature for everyday tasks: temperature 0.1, top_k 50, and a repetition penalty of 1.1. For math and STEM questions, it suggests temperature 1.0, top_p 0.95, and top_k 20.
Reasoning is on by default. The card warns that switching it off with enable_thinking=False can be inconsistent on reasoning-heavy tasks, so leave it on for anything tricky.
What it’s good at, and what to skip
Liquid AI points this model at local agents, tool calling, structured outputs, and fine-tuning for a narrow job. Think of a private assistant that fills forms, extracts data from text, or calls a few tools on your laptop.
The model card is blunt about the limits. It says the model is “not recommended for agentic coding or complex multi-step reasoning due to its small size.” Knowledge-heavy questions also need search or retrieval, because a 1.7B model can’t store much trivia.
A practical tip: start with a short, clear prompt and ask for a fixed format, such as a list of fields or a small JSON object. Small models follow tight instructions much better than open-ended ones, and the model card names structured outputs as a core use.
If you want a bigger local setup, our guide to running Qwen3.8-27B with Foundry Local covers the heavier end. For phone-side AI, see how to switch on Gemini Nano 4.
Frequently Asked Questions
Is Liquid AI LFM3 free to use?
The weights are free to download from Hugging Face. They ship under the LFM Open License v1.0, so read it before using the model in a commercial product.
How much RAM does LFM3-1.7B need?
About 1 GB at 4-bit precision, according to AI Weekly. Add some headroom for long prompts, since a 256K context window uses more memory as it fills.
Can LFM3-1.7B replace ChatGPT or Gemini?
No. It’s a small local model for focused tasks. It’s not meant for complex reasoning or agentic coding, and the model card says so directly.
Does it work offline?
Yes. Once the model file is on your device, a local runner like LM Studio or llama.cpp doesn’t need an internet connection.
Our take
If you’ve never run a model locally, LFM3-1.7B is a good first download. It’s small, fast, and free to try, and you’ll learn quickly what on-device AI can and can’t do. Don’t expect it to match a cloud chatbot, and don’t plan a coding workflow around it. Use it for private, repetitive text jobs and see how far it gets.
Leave a Reply