Liquid AI Unveils LFM3: Six Open-Weight 2B-Parameter Multimodal Models

What happened
Liquid AI introduced LFM3, a new family of six open-weight models with 2 billion parameters each. The models were released in June 2024 and are designed for a range of multimodal tasks, including text, vision, audio, and decision-making. All six models share a common architecture and can be fine-tuned on a single GPU. The models are available for download and use on Hugging Face and Liquid AI’s own platforms, with support for deployment across different devices.
Key details
- Each model has between 1.7 and 2.2 billion parameters.
- LFM3-1.7B supports a 256,000-token context window and fits in about 1 GB of memory at 4-bit precision.
- The family includes specialized models for text, vision (with video input), audio (supporting 18 languages), and combined modalities.
- Weights and documentation are available on Hugging Face and Liquid AI Playground.
- Commercial use is free up to $10 million in annual revenue under the LFM Open License v1.0.
Why it matters
Developers and companies now have access to a set of open-weight, multimodal models that can be fine-tuned and deployed on a wide range of hardware, including phones and laptops. This reduces the cost and technical barrier for organizations needing advanced AI capabilities for text, vision, audio, and decision tasks. The models’ open licensing and compatibility with popular AI platforms enable broader adoption and experimentation, especially for smaller businesses and independent developers.


