Liquid AI LFM3: what it is and how to run it

Liquid AI LFM3 is a new family of small open-weight models, each around 2 billion parameters, that cover text, vision, audio and yes/no-style decisions and are built to run on your own laptop, phone or server. You run it by downloading the weights from Hugging Face and loading them in Transformers, llama.cpp, Ollama, LM Studio or MLX. The text model fits in about 1 GB at 4-bit precision.

Liquid announced it on October 27. The @liquidai launch thread calls LFM3 "the ideal model suite for specialized intelligence" and lists six models across text, audio, vision and omni, plus a new tri-hybrid architecture tuned for 14 classes of GPUs, CPUs and NPUs. When we checked, the thread was close to 900,000 views. A later post in the same thread from @liquidai covers the vision model, which it says averages 71.4 across 62 vision benchmarks, the same score it gives Gemma 4 12B, at roughly a sixth of the size.

What's in the LFM3 family

From Liquid's Hugging Face release post:

The d1 models replace d1-3B, which we covered when it launched. Liquid says d1-2B scores 75.5 on its decision benchmark versus 64.0 for the older model, with about a third fewer parameters.

Under the hood, each model mixes gated short convolutions with two kinds of attention (sliding-window and global). Liquid trained the text model on 22 trillion tokens, and the last 3.5 trillion were agent traces and tool calls, which explains why the launch leans so hard on agent benchmarks.

The benchmark claim worth checking

The number getting attention is browser-agent performance. Liquid reports LFM3-1.7B scoring 38.42 on WebArena-Verified, against 3.02 for its own previous 2.6B model. VentureBeat, citing Liquid, puts Qwen3.5-9B at 28.94 on the same test, so a model a fifth the size comes out ahead.

Keep two things in mind. These are Liquid's own numbers, and VentureBeat notes nobody has independently verified them yet. And Liquid's model card is upfront that the text model is "not recommended for agentic coding or complex multi-step reasoning due to its small size." It's meant for narrow jobs you repeat a thousand times a day.

How to run LFM3

The quickest path is Python with Transformers:

pip install -U transformers

from transformers import pipeline

generator = pipeline("text-generation", model="LiquidAI/LFM3-1.7B", trust_remote_code=True)
messages = [{"role": "user", "content": "Pull the invoice number, date and total from this text: ..."}]
output = generator(messages, do_sample=True, temperature=0.1, max_new_tokens=512)
print(output[0]["generated_text"][-1]["content"])

If you don't write Python:

• In LM Studio, search for LFM3. The vision model is a 1.10 GB download.
• For Ollama or llama.cpp, use the GGUF build at LiquidAI/LFM3-1.7B-GGUF on Hugging Face.
• On an Apple Silicon Mac, there's a 4-bit MLX build, LiquidAI/LFM3-1.7B-MLX-4bit.
• If you'd rather not host anything, VentureBeat reports Liquid's console lists LFM3-1.7B at $0.08 per million input tokens and $0.20 per million output tokens, with context capped at 128K instead of 256K.

The vision and audio models need Transformers 5.20 or later, and their model cards have working examples.

Fine-tuning on one GPU

Liquid says every model in the family fine-tunes on a single GPU, and its cookbook on GitHub has Colab notebooks for real tasks: receipt photos to JSON, structured extraction from documents, spoken smart-home commands to function calls, and answering from retrieved documents or declining. That list reads a lot like the back-office work a small business would hand to a model like this.

Is LFM3 free?

The weights are free to download under the LFM Open License v1.0. Commercial use is covered only if your company, counted together with any entities under common control, makes under $10 million a year. Above that, you need a separate deal with Liquid. Qualifying nonprofits get an exemption for non-commercial and research use.

My read

A couple of years ago, a 2B model was a curiosity. LFM3's pitch is that you don't need a frontier model to read a receipt or sort a support ticket, and for jobs that narrow I think that's mostly right. Running it locally also keeps the data on your machine, the same appeal as Google's on-device Foresight note-taker.

I'm wary of the benchmark table, though, as I am with any self-reported scores. A model can top WebArena and still trip over your scanned invoices with coffee stains on them. Pick one boring, high-volume task, run it on 50 real examples, and count the mistakes before you trust it. If it holds up, that task stops costing you per-token API fees.

For businesses, the lesson here is that the biggest model isn't always the right one. A lot of everyday automation is pulling fields out of documents and deciding where things go, and that work is moving to cheap models you can run yourself. If you're not sure which of your processes fit, New Face Design's free process audit is a good place to start.