Liquid AI LFM3 Launch Brings Small Multimodal Models to Local Devices
The Liquid AI LFM3 launch introduces compact, open-weight AI models for specialized work. They can run on local hardware or hosted infrastructure. Liquid AI announced the release on October 27, 2026. Text, vision, audio, multimodal and decision-model variants share a common foundation.
The launch matters most to developers with a narrowly defined task. For example, developers can extract document fields, classify requests or interpret an image. These jobs may not need every input sent to a large remote model. In addition, the weights are available through Hugging Face, while Liquid also offers its API and Playground.
What does the Liquid AI LFM3 launch include?
Liquid says the first LFM3 release includes six models in the roughly 2-billion-parameter class. The text model, LFM3-1.7B, has 1.73 billion parameters. Its combined input-and-output context window is 256,000 tokens, according to the official model documentation. It is designed for local agents, tool calling, structured outputs and fine-tuning for a specific domain.
The wider family adds models for vision, audio, combined audio-vision-text input, and structured decision tasks. Liquid describes the decision models as tools for returning choices, scores, extracted spans or locations rather than simply generating chat responses. For instance, those formats could be useful for controlled workflows such as ticket routing or reviewing document fields.
Availability and the local-AI trade-off
Liquid says the models are available now as open weights and through its services in its launch announcement. The company also says the four-bit version of its text model is about 1 GB. That figure describes the model weights, not the complete memory required by an app: long context, supporting software and other runtime needs add to the total.
Running a model locally can give a team more control over where inference happens. It does not by itself make an application private or secure. A connected agent can still send data to search, messaging or other tools, so teams need to review the full workflow rather than only the model location.
Small models are not a replacement for every job
Liquid’s published benchmarks are company results, not independent proof that LFM3 will be best for a particular workload. VentureBeat’s review of the release notes mixed results. The text model led selected browser-automation and function-calling tests. However, it trailed larger rivals on several coding, knowledge and long-context evaluations.
The official documentation is similarly direct: LFM3-1.7B is not recommended for agentic coding or complex multi-step reasoning. As a result, a small pilot with clear accuracy, latency and human-review rules is a realistic starting point. It can show whether a general-purpose model needs replacing.
For more AI updates, visit Dumb IT Dude’s News archive.
Archive photograph of a laptop with technology and activism stickers, photographed by Leo Jiménez. Photo: Leo Jiménez (@leojimenezcr) / CC BY 4.0. Resized/display crop; original image license applies.
