Liquid AI released LFM3 on October 27, 2026, a family of six open-weight models in the 2-billion-parameter class for text, vision, audio, combined audio and vision ("omni") and two decision models, the company announced. All six branch from one shared backbone, LFM3-1.7B-Base, and are available on Hugging Face, the Liquid AI Playground and the Liquid AI API.
The text model, LFM3-1.7B, has a 256K-token context window and fits in about 1 GB at 4-bit precision. Liquid AI says it beats its previous LFM2.5-2.6B while using 35% fewer parameters. On WebArena-Verified, a benchmark of long web tasks, it scores 38.42, against 3.02 for LFM2.5-2.6B and 28.94 for Qwen3.5-9B. The company is clear that larger models still win on agentic coding and knowledge: on SWE-bench Verified it scores 43.40 against 51.20 for Qwen3.5-9B.
The architecture mixes three operators. Of 24 layers, 13 are gated short convolutions, 5 use sliding-window attention and 6 use global attention. The base model was pre-trained on 22 trillion tokens. Liquid AI tuned the design with hardware in the loop across 14 device and backend combinations, including two phones, a Raspberry Pi 5, laptops, cloud CPUs, GPUs and a Qualcomm NPU.
Every model can be fine-tuned on a single GPU. The company's example: a few hours of training on one A100 raised the vision model's accuracy at classifying wind turbine blade defects to 67%, from under 10% out of the box. MacPaw, a launch tester, reports that LFM3-1.7B beat Qwen3.5-2B on 65 everyday agent tasks on macOS, with 64% task success against 48% and 58% fewer tokens per task.
All benchmark figures are Liquid AI's or its partners'. Day-zero support includes llama.cpp, MLX and vLLM, with launch partners including AMD, Arm, NVIDIA and Qualcomm.
Why this matters
A capable agent model in about 1 GB changes where AI can run: on a phone, a laptop or a server inside a company's own network, with no data leaving the device. For regulated teams, the useful test is not the leaderboard. It is whether a small model, fine-tuned on their own data, handles one narrow task reliably enough to stay on premises.
This article is not legal advice.
