Six small models, one backbone: Liquid AI's LFM3 puts agents, vision and voice in about 2B parameters 💧

On 27 October, Liquid AI released LFM3, a new family of open-weight models on Hugging Face. The idea is simple and, to me, very appealing: instead of one huge general model, you get six small ones (text, vision, audio, omni and two "decision" models), all trained from the same 1.7B-parameter base. You pick the one that matches your inputs, fine-tune it on a single GPU, and run it close to where your app lives.

What's actually new

The base model, LFM3-1.7B-Base, uses what Liquid calls a tri-hybrid layer mix: 13 gated short convolutions, 5 sliding-window attention layers with a 256-token window, and 6 global attention layers. Each attention head has a sigmoid gate, so a head can effectively switch off when it has nothing to add, which the team says can help with quantization. It was pre-trained on 22T tokens, and the last 3.5T tokens lean more on agentic traces and tool calls, which also stretches the context window to 256K tokens.

Liquid didn't design this on paper alone. Their hardware-in-the-loop architecture search now covers 14 device and backend combinations, from phones and a Raspberry Pi 5 to laptops, cloud CPUs, GPUs and the Qualcomm Hexagon NPU.

Why I find this exciting

The numbers that caught my eye are about doing more with less. According to Liquid's post, LFM3-1.7B beats its predecessor LFM2.5-2.6B with 35% fewer parameters, and on the WebArena-Verified agent benchmark it scores 38.42 against 3.02. The text model fits in about 1 GB at 4-bit precision, and it is post-trained to run inside local agent harnesses such as Hermes Agent, Pi and OpenClaw.

The vision model, LFM3-VL-2B, is the first LFM vision model that accepts video, and Liquid reports its average over 62 vision benchmarks rising from 57.6 to 71.4 compared with LFM2.5-VL-3B. The audio side is just as practical: a fully causal audio encoder that can start processing while you're still speaking (time to first token as low as 50 ms), and a speech codec that runs at about 87x real time on a single CPU thread on an Apple M4 Max.

What I like most is the attitude. The blog ships ready-made fine-tuning notebooks (SFT, DPO, GRPO and continued pre-training), from extracting JSON out of messy documents to calling the right tool among 108 in a phone app. For anyone running a homelab or building focused tools, like a tutor that helps students with one subject, that's exactly the kind of starting point that turns "maybe someday" into a weekend project.

The honest caveats

These are the developer's own benchmarks, so independent tests will tell the fuller story. The models are open-weight but not Apache-licensed: the model card lists Liquid's own LFM Open License v1.0, and PacketNebula's tracker notes that this licence limits commercial use to organisations under $10M in annual revenue. And a 2B model is still a 2B model: it shines when specialised, not as a replacement for frontier generalists.

Sources

Small, fast and built to be specialised: I'm curious to try LFM3-1.7B as a local agent on my own hardware.

Reacties

Een reactie posten

Populaire posts van deze blog

Meta just shared the full recipe for a 30B agent model you can run at home 🧠

When Meta released Muse Glimmer in August, the pitch was simple: an open-weight, Apache 2.0 model built for AI agents that run on your own Mac or PC. What was missing was the "how". On 19 October, Meta Superintelligence Labs published the full technical report, and for anyone who cares about distillation and local agents, it is one of the most useful reads in a while. What the model is Muse Glimmer is a dense model of about 29.6 billion parameters, including a 1.8B perception encoder, so it reads text and images and writes text. It has a 128K+ context window and was trained on data from more than a hundred languages. The design goal was strict: everything has to fit in 32 GB of GPU memory, and a smaller 4-bit variant targets 24 GB cards. The parts I found most interesting Dense beat mixture-of-experts here. That surprised me. The team trained scaling ladders for a dense model and two MoE designs with the same memory footprint. Under a fixed memory ceiling and their compu...
Meer lezen

AI that can actually "see" a crystal 🔬

Most of the solid stuff around us, from battery materials to chips, is made of crystals: atoms stacked in neat, repeating 3D patterns. Large language models are great at reading papers about those materials, but they have never really looked at the structures themselves. A team at MIT just showed a clever way around that, and the trick is one I keep coming back to: good synthetic data. What they built The model is called CLaC, short for Contrastive Language-Crystal model, and it is described in a paper in Nature Communications by Yang Jeong Park, Mayank Kumaran, Chia-Wei Hsu, Elsa Olivetti and Ju Li. MIT News covered it on 18 October 2026. If you know CLIP, the model that taught image generators to connect pictures and words, you already get the idea. CLaC has two branches: a graph neural network that reads the 3D crystal structure, and a text encoder that reads a description. Contrastive training pulls matching structure-text pairs together in one shared space and pushes mismatche...
Meer lezen

A 35B model on a Mac mini, using just 2.9 GiB of memory 🚀

There's something special about running a capable language model on hardware you own: no API bill, no data leaving the house, just a small box on your desk doing the thinking. A paper that's currently #2 on Hugging Face's trending list takes that idea further than I expected: it runs a 35-billion-parameter model on a Mac mini at about 20 tokens per second while using under 3 GiB of memory. The trick: leave the experts on the SSD The paper is called The Other Half of the Memory Wall: Serving 35B MoEs from SSD with Trained Routing Prediction , and the system is named Edge0 . It builds on a Mixture-of-Experts (MoE) model, Qwen3.6-35B-A3B. In an MoE, each token only uses a few "experts" per layer (here 4 out of 256 in each of 40 layers), so most of the weights sit idle at any given moment. Edge0 takes that literally. The expert weights (quantized to int4) stay on the SSD and are streamed in only when needed. The clever part is a small prerouter : a tiny trained he...
Meer lezen