LFM3: Features, Capabilities and How to Use It

What is LFM3?

Liquid AI developed LFM3 as a new generation of Liquid Foundation Models built for specialized local intelligence. The 2B-parameter model family shares a unified backbone and provides open weights for text, vision, audio, omni, and decision tasks.

The foundation models are trained on 22 trillion tokens to provide higher expressivity while running directly on consumer and edge devices. Developers can fine-tune every model on a single GPU, download open weights from Hugging Face, or run requests using the Liquid AI API.

How does the LFM3 architecture work?

Liquid AI designed LFM3 with a 24-layer tri-hybrid architecture that combines gated short convolutions, sliding-window attention, and global attention. This combination maintains fast context processing on resource-constrained devices without sacrificing quality.

Of its 24 layers, 13 are gated short convolutions with a kernel size of 7, 5 are sliding-window attention layers with a 256-token window, and 6 are global attention layers. The design incorporates head-wise gated attention to suppress activation outliers and partial RoPE to stabilize long-context processing.

To optimize performance, Liquid AI tested LFM3 across 14 hardware combinations including mobile phones, a Raspberry Pi 5, laptops, and cloud processors. With an 8,000-token context, the LFM3-1.7B backbone runs about 25 percent faster than LFM2-1.2B while matching its short-prompt performance per parameter.

Which models are included in the LFM3 family?

The LFM3 release features six specialized models that all branch from the base LFM3-1.7B foundation. Developers can select a model tailored to specific input modalities rather than deploying a single large model.

The text model is post-trained to run in agent harnesses like Hermes Agent, Pi, and OpenClaw. The vision and audio models include custom in-house encoders, such as a 164M visual encoder that uses 2.2 times fewer FLOPs than Qwen3.5-2B's vision encoder.

What devices and platforms support LFM3?

Liquid AI optimized LFM3 to run across laptops, mobile phones, edge chips, and cloud environments. Supported systems include hardware powered by AMD Ryzen AI, Snapdragon, Apple Silicon, and Ambarella silicon.

On-device speech generation uses LiquidCodec, an audio tokenizer that generates speech at approximately 87 times real-time on a single Apple M4 Max CPU thread. The text model requires about 1 GB of memory at 4-bit precision, making it suitable for continuous background background tasks.

How to access and run LFM3

Developers can download LFM3 open weights from Hugging Face or access the models via the Liquid AI Playground and API. Running LFM3 locally requires Python and the Hugging Face Transformers library.

1. Install or update the Transformers library using pip install -U transformers.
2. Initialize the text model pipeline using pipeline("text-generation", model="LiquidAI/LFM3-1.7B", trust_remote_code=True).
3. Pass user prompt dictionaries into the generator pipeline to receive generated text.
4. To process visual inputs, install transformers>=5.21.0, torch, torchvision, and pillow.
5. Load LiquidAI/LFM3-VL-2B using AutoModelForImageTextToText and pass image URLs or tensors using AutoProcessor.

What are the limits of LFM3?

While LFM3 delivers high efficiency on edge hardware, the initial release is limited to small 2B-class parameter sizes. Tasks requiring vast parameter scale or deep multi-step general knowledge may still require larger cloud-hosted models.

To run models with custom toolsets or specialized domain knowledge, developers must perform local fine-tuning. Additionally, running multimodal models locally requires installing specific dependencies like PyTorch, Torchvision, and Pillow.

Frequently asked questions

Sources: Liquid AI · Hugging Face