LFM3-VL-2B
LFM3-VL-2B is a multimodal vision-language model developed by Liquid AI as part of the LFM3 family, designed specifically for efficient, on-device visual understanding and spatial reasoning. Released in October 2026, the model combines a compact language backbone with an in-house hybrid vision encoder to process text, images, and native-resolution video.
Architecture and Design
The model features a total parameter count of approximately 1.92 billion, broken down into a 1.73B language model, a 164M vision encoder, and a 25M projector. The language model backbone uses a tri-hybrid architecture consisting of 24 layers: 13 gated short convolutions, 5 sliding-window attention layers, and 6 global attention layers. The vision encoder incorporates gated depthwise 2D convolutions and bidirectional attention processing, operating on 16x16 patches with 2D rotary position embeddings.
Key Capabilities
LFM3-VL-2B supports a context length of 131,072 tokens and handles text, image, and video input modalities. Key use cases include answering questions about recorded video feeds, parsing complex documents, charts, and application screens, and providing spatial reasoning for robotics applications. It also incorporates step-by-step reasoning capabilities to improve accuracy on multi-step visual problem-solving tasks.
Create with Crafiq
Generate images, 3D models, video and audio in one studio.
Explore the studioHow LFM3-VL-2B ranks
LFM3-VL-2B is highlighted in the table below. Switch the metric to see how the ordering changes.