Recently, the renowned AI innovation company Liquid AI officially unveiled its next-generation foundation model family—LFM3. The initial release comprises six models with parameter sizes around 2 billion, comprehensively covering multiple domains including text, vision, audio, omni-modal, and decision-making. All are based on the LFM3-1.7B-Base foundational architecture and achieved a staggering 22 trillion token throughput during pre-training.
A groundbreaking tri-hybrid architecture: fewer parameters, greater efficiency
LFM3 employs a brand-new tri-hybrid layer structure, seamlessly integrating gated short convolutions, sliding-window attention (SWA), and global attention mechanisms. Across its total 24-layer architecture, there are 13 convolutional layers, 5 sliding-window attention layers, and 6 global attention layers.
Compared to the previous generation, LFM3 features a complete architectural overhaul and significantly expanded hardware compatibility testing, covering 14 different device and backend combinations—including smartphones, Raspberry Pi, laptops, and various cloud-based CPUs, GPUs, and Qualcomm NPUs. Testing demonstrates that, thanks to its innovative design, LFM3-1.7B not only maintains consistent performance with its predecessor on short prompts but also boosts processing speed by approximately 25% when handling 8K-context inputs. Meanwhile, its text model achieves a size of just about 1 GB even under 4-bit quantization, greatly lowering the barrier for edge-side deployment.
Multimodal and decision-making capabilities spanning all scenarios
The six newly released 2-billion-parameter models each boast unique strengths, perfectly suited to diverse application needs:
Text model (LFM3-1.7B): It offers a context window of up to 25K and has undergone extensive post-training for agent workflows. It performs strongly on instruction following and function calling across multiple benchmarks, runs smoothly on a single GPU, and supports coordination among large numbers of background agents.
Vision model (LFM3-VL-2B): It introduces video input support and enables step-by-step reasoning, spatial visual reasoning, and graphical user interface understanding. With a newly developed in-house vision encoder, its computational efficiency improves substantially over the previous generation.
Audio model (LFM3-Audio-2B): It supports speech recognition in 18 languages, multilingual speech synthesis, audio understanding, and real-time voice conversations. Its fully causal architecture reduces first-token latency to as low as 50 milliseconds.
Omni-modal model (LFM3-Omni-2B): It brings vision, audio, and language together in a unified model, supporting spoken questions and answers, audiovisual video understanding, and cross-modal tool calls for more natural human-computer interaction.
Decision models (d1-2B and d1-Omni-2B): Designed for rapid millisecond-level decision-making, they directly produce precise structured decisions from complex text, image, or audio inputs, eliminating cumbersome interactions with cloud services.
Enabling specialized industries and on-device robotics
Beyond its broad general capabilities, the LFM3 family emphasizes specialized intelligence and easy fine-tuning from the outset. Developers can fine-tune models at low cost on 1 GPU for specialized industrial tasks such as detecting defects in wind turbine blades. Through LoRA adaptation and end-to-end training, LFM3-VL-2B also supports robotic manipulation, visual navigation, and multimodal embodied control, enabling efficient real-time inference on edge hardware such as Jetson.
The open weights of the LFM3 series are now available, with release-day support from AMD, Qualcomm, NVIDIA, Apple MLX, and major development frameworks and platforms.