JOURNAL / 2026.10.28

Liquid AI launches LFM3: local vision and voice with different limits for each model

The new family publishes weights for small text, vision, audio and decision models. The opportunity lies in specialized local applications; context windows and runtime support require reading each model card separately.

Liquid AI introduced LFM3 on October 27. The family shares a backbone and branches into six models with roughly two billion parameters: text, vision, audio, a multimodal combination and two decision variants. The weights are already published. The practical attraction is choosing and adapting a small component for an application that needs to work close to its data.

The architecture introduces sliding-window attention alongside the short convolutions and global attention of the previous generation. Liquid reports 22 trillion pretraining tokens and makes specialization a central goal. This gives the release a technical direction, but neither training volume nor a shared backbone makes the variants interchangeable products.

The LFM3-1.7B model card describes a text model for agents, tools and structured outputs, with a 256K-token context window. In vendor tests it improves over LFM2.5-2.6B on BFCL v4, which evaluates function calls, and ToolSandbox, which adds interactions involving tools and state. This is relevant evidence for automation; it does not establish sustained reliability in a particular application. The card itself advises against agentic coding and complex multi-step reasoning.

The LFM3-VL-2B vision variant introduces video input to LFM vision models. It supports 128K tokens and samples video at two frames per second. It could help answer questions about a recording or locate elements in an image, but that sampling calls for checking what happens to brief events. Its card also advises against multi-turn conversations and complex reasoning.

LFM3-Omni-2B combines images, video, audio and text, and responds with text or speech. It has approximately 2.17 billion parameters, a 32K window and no reasoning mode. The distinction matters: the text model’s 256K does not describe the conversational capacity of the entire family.

Consider a maintenance application receiving a photo and the spoken question “Which part am I pointing at?” Omni makes that combined interaction worth testing; a vision variant could fit if the question arrives in writing. My reading is that the potential gain lies in reducing components and adapting the system to a specific vocabulary. We would still need to check whether it identifies the correct part among nearly identical models and asks for another photo when evidence is missing. This is a proposed use, not an observed result.

There is also a distance between publishing weights and having every runtime path mature. Although the announcement highlights support from launch, the vision card says its llama.cpp integration remains under review and does not yet include video; it also notes pending work in vLLM and SGLang. The choice depends on the model, modality and runtime version. A startup example does not guarantee that a stable installation supports all those functions.

Memory raises a similar issue. Liquid places the text model at around 1 GB with four-bit quantization, but its device speed comparisons use 32K context. That initial figure is not a complete budget for a 256K conversation, a camera and a voice interface. I would measure peak memory, time to first response and consumption during a representative session, alongside errors requiring intervention.

Downloadable weights do not remove usage conditions either. The LFM Open License 1.0 includes a commercial restriction tied to an annual revenue threshold of $10 million. Its terms need reviewing for the particular deployment.

LFM3 deserves attention because it broadens what a team can try to specialize and keep under its control. For me, the decisive test would be a narrow task that works better by combining perception and language, with a reproducible runtime path and identifiable errors. That will show how much of this release becomes a useful application.

Primary sources consulted on October 28, 2026. We reviewed model cards and weight-file availability; we have not run the models or reproduced the evaluations.

← Back to journal