Recently, the well-known AI innovation company Liquid AI officially released a new generation of basic model family — LFM3. The first version of this series includes six models with parameter sizes around 2B. They cover multiple dimensions such as text, vision, audio, omni-mode, and decision-making. All of them are based on the LFM3-1.7B-Base infrastructure, and they processed up to 22T tokens during the pre-training phase.
Breakthrough 3 Hybrid Architecture: Fewer parameters, higher efficiency
LFM3 adopts a completely new Tri-hybrid layer mix structure, organically integrating gate short convolution, sliding window attention mechanism (SWA), and global attention. In the total of 24 layers, there are 13 convolutional layers, 5 sliding window attention layers, and 6 global attention layers.
Compared to the previous generation of products, LFM3 has undergone a comprehensive rearchitecture at the architecture level, and its hardware adaptation testing scope has been significantly expanded. It covers 14 different device and backend combinations, including smartphones, Raspberry Pi devices, laptops, as well as various cloud-based CPUs, GPUs, and Qualcomm NPUs. Testing shows that with its new design, LFM3-1.7B not only maintains the same performance for short prompts as the previous generation, but also increases speed by approximately 25% when handling 8K contexts. Meanwhile, its text model occupies only about 1GB in 4-bit quantization, greatly reducing the barrier to on-premise deployment.
The multimodal and decision matrix that covers all scenarios
The six newly released 2B-class models each have unique features, perfectly adapting to various application scenarios:
Text Model (LFM3-1.7B): It has a context window of up to 25K and has undergone deep post-training for Agent workflows. In multiple benchmark tests, its instruction following and function calling capabilities are outstanding. It can run smoothly on a single card and supports extensive collaboration among background agents.
Visual Model (LFM3-VL-2B): First to introduce video input capabilities, supporting step-by-step deep thinking, spatial visual reasoning, and graphical interface understanding. With the new self-developed visual encoder, its computational efficiency has been significantly improved compared to the previous generation.
Audio model (LFM3-Audio-2B): Supports speech recognition in 18 languages, multilingual synthesis, audio understanding, and real-time voice chat. It adopts a fully causal architecture, with a first-word latency as low as 50 milliseconds.
Full Modal Model (LFM3-Omni-2B): For the first time, vision, audio, and language are integrated into a single model. It supports voice-question answering, audiovisual video understanding, and cross-modal tool invocation, providing a more natural human-machine interaction experience.
Decision-making models (d1-2B and d1-Omni-2B): Specifically designed for millisecond-level rapid decisions. They can directly output precise, structured decision results for complex text, images, or audio inputs, eliminating the need for cumbersome cloud-based interactions.
Empowering vertical fields and edge robots
In addition to its powerful general capabilities, the LFM3 family emphasizes “specialized intelligence” and easy fine-tuning from the design stage. Developers can perform low-cost fine-tuning of models using a single GPU, enabling its application in vertical industrial fields such as wind turbine blade defect detection. Furthermore, through LoRA adaptation and end-to-end training, LFM3-VL-2B can seamlessly integrate into robot operations, visual navigation, and multi-modal embodied intelligence control, achieving efficient real-time inference on edge hardware like Jetson.
Currently, the open-source weights for the LFM3 series models have been officially released, and they have received support from AMD, Qualcomm, NVIDIA, Apple MLX, as well as major mainstream development frameworks and platforms on the first day.