Liquid AI releases LFM3 for local agents and multimodal apps
Liquid AI released LFM3 on October 27, 2026. The open weight family includes six models around the two billion parameter size class, covering text, vision, audio, combined modalities and structured decisions.
The useful starting point is the job your application needs to finish. A model that can produce a spoken answer solves a different problem from one that extracts a field or chooses between predefined options. Pick that requirement before comparing headline scores.
Choose the inputs and outputs first
| Checkpoint | Inputs | Outputs |
|---|---|---|
| LFM3-1.7B | Text | Text |
| LFM3-VL-2B | Text, images, video | Text |
| LFM3-Audio-2B | Text, audio | Text, speech |
| LFM3-Omni-2B | Text, images, video, audio | Text, speech |
| d1-2B | Text, images, video | Calibrated answers |
| d1-Omni-2B | Text, images, video, audio | Calibrated answers |
For a new project, write the input and required response on one line. For example, a receipt photo in and invoice fields out. Then decide whether the application really needs conversation, speech or several input types. This keeps optional capabilities from becoming accidental deployment requirements.
What the small text model actually promises
The LFM3 text model card lists 1.73 billion parameters and a 256K token context covering input and output together. Liquid describes about 1 GB of memory at four bit precision. That headline should not be treated as a guarantee for an entire application at maximum context.
The card also sets a useful boundary. It recommends the model for local agents, tool calls, structured outputs and specialization, while advising against agentic coding and complex reasoning across many steps. For knowledge intensive tasks, it recommends retrieval or search.
Measure the configuration you intend to ship. Include your actual prompt length, conversation history and simultaneous requests. Report peak process memory and response time together, rather than using the download size as the hardware requirement.
Check runtime support before downloading
Liquid publishes a GGUF version for local inference. Its instructions require a llama.cpp build with LFM3 support. The repository also distinguishes a Q4_0 checkpoint trained through quantization aware distillation to recover accuracy lost during quantization.
Record the exact checkpoint, quantization and runtime version in your trial notes. If two tests use different files or engines, do not attribute every difference to the model family. Reproduce a small reference request before adding application logic.
The commercial license needs a separate check
The text model uses the LFM Open License v1.0. Liquid’s license page describes a $10 million annual revenue threshold for free commercial use and directs larger commercial users to arrange a separate license. It also requires retaining notices when redistributing models or derivatives. Downloadable weights should not be read as unrestricted Apache 2.0 permission.
What Liquid’s own demos actually show
Liquid’s launch gallery includes a flight booking demonstration in which LFM3-Omni combines spoken input, screen understanding and structured tool calls. A separate event flyer demonstration connects information captured on camera with a calendar action. These are vendor examples, not independent reliability measurements.
The d1 object finder demonstration illustrates a different job. On a MacBook, it draws boxes around objects matching a plain language description in one forward pass. The useful output is a location, rather than a conversational answer.
For a calendar assistant, correct transcription is only one requirement. An ambiguous date, duplicate event or missing time zone can still produce a bad action. Before connecting real accounts, test those cases with mock tools and require confirmation for consequential changes. This is our suggested acceptance test, not a test we performed.
