Liquid AI releases small models for local text, vision and voice tasks

Liquid AI announced LFM3 on October 27, releasing an open-weight model family built around a small shared backbone. The launch covers text, vision, audio, combined multimodal input and decision models in roughly the two-billion-parameter range. The weights are available through the company’s Hugging Face collection, giving developers something they can download and evaluate now. Liquid’s announcement · Official collection

The practical stake is where useful AI runs. A model small enough for a laptop or phone could support tasks such as organizing documents, interpreting an image or responding to a spoken request without sending every input to a remote model. These are building blocks for applications, rather than a complete personal assistant that becomes reliable simply by downloading a file.

Several tools built from a small backbone

The family includes LFM3-1.7B for text, LFM3-VL-2B for vision, LFM3-Audio-2B for sound and speech, and LFM3-Omni-2B for combined vision, audio and language. Liquid also lists d1-2B and d1-Omni-2B decision models, alongside the base model. Developers can select a model for a narrower task instead of running every capability in one larger system. Release overview

The text model card specifies 1.73 billion parameters and a maximum context window of 256,000 tokens across input and output. Parameters are the learned values in a model; fewer of them can reduce the resources needed to run it. Context is the material a model can consider within a request. A large advertised context ceiling does not guarantee that every detail will be recalled correctly or that the longest request will be practical on every device. Text model card

Liquid says the text model’s weights take about 1 GB at four-bit precision. That is a compressed model-footprint claim, not a promise that the whole application needs only 1 GB of RAM. Runtime software, working memory, context and other components add to the total. The company’s own launch discussion reports additional cache memory for long contexts. Hardware fit needs to be measured with the actual model format, application and workload.

Local processing offers a choice, not automatic privacy

A business could evaluate a model for extracting structured fields from a familiar type of document. A developer could explore voice commands or image questions within an application they control. Liquid supplies local inference options and describes fine-tuning on a single GPU to specialize the models. Bright has not run these models or verified their speed, quality or fine-tuning requirements. Model formats and guidance

Keeping inference on a device can reduce how much content needs to leave it. Privacy still depends on the surrounding application: search tools, cloud storage, telemetry and external services may transmit information. A local language model connected to an online tool does not make the tool’s activity local. Teams need to check the full data path before promising private or offline operation.

That distinction is especially relevant for agents, software that uses a model to choose and call tools. A model can propose an action, but the application supplies its permissions and executes it. Downloadable weights do not settle whether an agent should be allowed to send messages, change records or handle sensitive files.

Read the benchmark claims with their boundaries

Liquid reports improvements over its earlier models and competitive results against larger models on selected evaluations. Those results are vendor-reported, with different tasks and setups represented in the launch tables. They support a reason to test the model for a particular job; they do not establish that it replaces a larger general-purpose system across workloads. Published evaluation tables

The text model card is more specific about intended use: local agents, tool calling, structured outputs and domain-specific fine-tuning. It recommends retrieval or search for knowledge-heavy work and discourages using the model for agentic coding or complex multi-step reasoning because of its small size. Those limits belong beside the performance claims when deciding whether to adopt it. Intended use and limitations

Downloadable does not mean unrestricted

LFM3-1.7B uses the custom LFM Open License v1.0. It is based on Apache 2.0 but adds a commercial-use condition tied to a $10 million annual-revenue threshold. Larger businesses need a separate commercial agreement with Liquid. Redistribution also carries license and attribution requirements, while qualified nonprofits have an exemption for non-commercial or research use. Readers planning commercial deployment should review the full terms for their circumstances. Liquid’s license

That makes “open-weight” the useful description here. The weights can be obtained, inspected and adapted subject to their license; the release is not simply under standard Apache 2.0 terms. Access to a downloadable model and unrestricted commercial permission are different questions.

For a first evaluation, the useful test is a bounded task with representative examples: does the model produce the required format, handle errors and uncertainty, and meet the device’s memory and response-time needs? LFM3 makes that experiment available now. Its contribution to everyday local AI will be measured by applications that work reliably, not by the small parameter count alone.