Liquid AI
Liquid AI’s LFM3 fits a 256K-context model into about 1 GB
Six open models of around 2 billion parameters, built to run on phones and laptops, with a text model that beats Qwen3.5-9B on Liquid’s benchmark average.
Liquid AI released LFM3 on 27 October 2026, a family of small open-weights models designed to run on phones, laptops and other devices instead of in a data centre. The first release has six models of around 2 billion parameters, all branched from one base model: text, vision, audio and omni models, plus two decision models derived from the vision and omni versions.
Small, with a long memory
The text model, LFM3-1.7B, fits in about 1 GB at 4-bit precision and supports a 256,000-token context window. Liquid says it outperforms its previous LFM2.5-2.6B with 35% fewer parameters. The architecture mixes gated short convolutions with sliding-window and global attention, which keeps long contexts cheap: a full 128,000-token context needs 1.6 GB of cache memory in llama.cpp, against 4.3 GB for Qwen3.5-4B. The base model was pre-trained on 22 trillion tokens.
Made to be specialized: every model can be fine-tuned on a single GPU.
Where it is strong, and where it is not
Across Liquid’s benchmark set, LFM3-1.7B averages 49.37, ahead of Qwen3.5-9B at 47.24 despite being a fraction of the size. It leads on instruction following and function calling. Liquid is clear that larger models still win on agentic coding and general knowledge: on GPQA Diamond it scores 63.37, against 80.40 for Qwen3.5-9B.
The weights are on Hugging Face under the LFM Open License v1.0. It allows commercial use by organisations with annual revenue under US$10 million; commercial use by larger organisations is not covered by the licence. The models are also available in Liquid’s playground and API.
What it means for your business
A model that runs on the device keeps the data on the device. For field staff, clinics, trades businesses and anyone handling personal information, that means some AI tasks, such as summarising notes, filling in a form from a photo or routing a request, can run without sending data offshore and without a connection.
The licence suits most Australian small and medium businesses, but check it if your revenue is near the US$10 million line. The bigger point is how work gets split: small local models can handle the private, routine steps and larger hosted models the hard ones, with your team deciding which is which.
Sources
- LFM3: Liquid Foundation Models for specialized intelligence (opens in a new tab) · Liquid AI
- LFM3-1.7B model card (opens in a new tab) · Liquid AI on Hugging Face
- LFM Open License v1.0 (opens in a new tab) · Liquid AI on Hugging Face
General information only, not legal, financial or professional advice. Check the sources for your circumstances.