Photo: Merlin Lightpainting / Pexels
Six open-weight multimodal models of roughly 2 billion parameters each are built to run and fine-tune on a single GPU
Liquid AI is betting that the next big thing in artificial intelligence is actually pretty small. On October 27, 2026, the company launched its LFM3 family: six open-weight multimodal models built to run and be fine-tuned on local hardware.
Each model carries approximately 2 billion parameters. In an industry where model sizes often get announced like stadium attendance figures, that is a deliberately modest number.
The pitch is simple. You should not need a data center to get useful AI, and you should be able to customize it on a single GPU.
Six models, one backbone #
All six models share a common foundation called LFM3-1.7B-Base. Think of it as one chassis with different bodies bolted on top, each tuned for a different kind of input.
The lineup spans text, vision-language, audio, decision-making, and an “omni” model. That omni variant handles text, images, video, and audio together, which makes it the Swiss Army knife of the bunch.
The text model, LFM3-1.7B, supports a context window of 256K tokens. A context window is how much material a model can hold in its head at once, so a larger one means it can work through long documents without losing the thread.
The vision-language model, LFM3-VL-2B, is where Liquid AI is making its boldest claim. According to the research findings, it has shown performance matching models more than six times its size on vision tasks.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Because the models are open-weight, developers can download them and adapt them to their own needs. Liquid AI is highlighting domain-specific uses such as robotics and defect detection, the sort of jobs where a model sits close to the action rather than in a distant cloud.
The tri-hybrid architecture #
The technical centerpiece is what Liquid AI calls a tri-hybrid architecture. It combines three components: gated short convolutions, sliding-window attention, and global attention.
Gated short convolutions handle the nearby, local patterns efficiently. Sliding-window attention looks at a moving chunk of recent context, while global attention lets the model connect ideas across the entire input.
The architecture was pre-trained on 22 trillion tokens. The result, according to the research findings, is a high performance-per-parameter ratio compared with larger models.
Who is Liquid AI #
Liquid AI was founded in 2023 and has raised approximately $287.5 million since then. The largest piece of that came from a $250 million Series A round in December 2024.
Edge hardware refers to devices that process data where it is created, such as robots, cameras, and factory equipment. Running AI on those devices directly can cut latency and keep sensitive data from ever leaving the building.
What this means #
For developers and enterprises, the most practical detail is the single-GPU fine-tuning. A manufacturer that wants a model to spot defects on its own production line does not necessarily need a general-purpose giant. It needs something small, cheap to adapt, and good at one specific job. That is precisely the niche LFM3 is targeting. The open-weight approach also matters. Teams can run the models on their own hardware, inspect them, and tailor them without routing every request through a third-party service.
There are caveats worth watching. Benchmark claims, including the six-times-larger comparison for LFM3-VL-2B, will be tested by independent developers once the models are in wider use, and real deployments often look messier than lab results.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our