This week: Liquid AI released LFM3, a family of six small open-weight models that read, see, listen, talk, and make quick decisions, all sized to run on a phone or laptop instead of a data center.

Elsewhere: Why a 1.7B model is suddenly doing agent work, what a few hours of fine-tuning did for wind turbine inspections, and where the big models still win.

The New Prime Signal

For a couple of years, the AI story has mostly been "bigger." Bigger models, bigger data centers, bigger power bills. Liquid AI keeps making the opposite bet, and on October 27 it announced LFM3: six models, each around 2 billion parameters, that handle text, images, video, speech, and fast yes-or-no style decisions. The text model "fits in about 1 GB at 4-bit precision," which is roughly the size of a couple of phone videos. Somewhere, a server rack is feeling a little underappreciated.

What happened: LFM3 is one shared backbone with six branches: a text model (LFM3-1.7B), a vision model that can now watch video (LFM3-VL-2B), an audio model that listens and talks back in 18 languages, an "omni" model that combines sight and sound, and two "decision" models that answer a list of questions about a ticket, invoice, photo, or voice note in one pass. Liquid says it pre-trained the base model on 22T tokens and tuned the design while testing on "14 device and backend combinations," including two phones, a Raspberry Pi 5, laptops, cloud CPUs, GPUs, and a Qualcomm NPU. The weights are on Hugging Face, with a browser playground and an API alongside. Launch partners include AMD, Arm, Intel, NVIDIA, and Qualcomm, and Liquid says Mercedes-Benz and Shopify helped test before release.

Why it matters: A model this small can run on the device in your hand, which means no cloud bill, no internet requirement, and your data stays put. Liquid's headline claim is that the new 1.7B text model can "outperform our previous LFM2.5-2.6B while using 35% fewer parameters," and on its own agent and tool-use tests it "competes with models several times its size". One outside tester backs part of that up: MacPaw, the Mac software company, says that on 65 everyday agent tasks on macOS, LFM3-1.7B hit "64% task success versus 48%" for Qwen3.5-2B while using 58% fewer tokens. Liquid also says it ran "more than 300,000 LFM3-1.7B agents in parallel" to dig through the Enron email archive, which is either a great stress test or the worst office party ever.

We like this one. Not every job needs a genius in the cloud; plenty just need a sharp assistant that lives on your laptop and doesn't phone home. The interesting idea isn't any single benchmark; it's the menu: pick the small model for your task, teach it your niche, and run it where the data already is. We'd still want independent tests before trusting a 1.7B model with anyone's inbox. But if the numbers hold, "small and specialized" just got a lot more practical.

Liquid's pitch is that you shouldn't use these models as-is. You should make one an expert in your thing.

How it works: Under the hood, LFM3 mixes three kinds of layers: fast "short convolutions" that look at nearby words, attention over a sliding window of recent text, and a few full attention layers that can look back over the whole conversation, up to a 256K-token context for the text model. Liquid says that mix keeps long chats fast and memory-light on small devices. Then comes the specializing. In one demo, Liquid fine-tuned the vision model to spot defects in drone photos of wind turbine blades, using "about 3,000 public photos" and "a few hours on one A100." Out of the box, both it and a big cloud model identified the defect types in "fewer than 10% of images." After that short fine-tune, classification accuracy rose to 67%. Liquid also showed the same vision model being adapted to steer a simulated robot arm and to navigate, though those are research demos, not products.

Try it: The quickest way in is the Liquid Playground in your browser. If you're comfortable with local AI tools, the LFM3-1.7B text model and the LFM3-VL-2B vision model are on Hugging Face, with native support for llama.cpp, MLX, and vLLM, and Liquid lists LM Studio and Ollama among its supporting tools. Liquid's docs include fine-tuning notebooks and say each model can be fine-tuned on a single GPU. Start with something boring and useful, like sorting your own receipts or tagging photos, and see if it holds up before handing it anything important.