Liquid AI has released LFM3, the latest generation of its Liquid Foundation Models, designed as starting points for specialized super intelligence (SI) applications. The new family brings text, vision, audio, omni, and decision models onto a shared 2B-parameter backbone, allowing developers to fine-tune SI systems for specific tasks on a single GPU and deploy them directly to edge devices.
What Happened
The LFM3 release introduces a tri-hybrid architecture that combines gated short convolutions, sliding-window attention, and global attention. According to the company, this design allows the LFM3-1.7B model to outperform its predecessor, LFM2.5-2.6B, while using 35% fewer parameters. The family includes six models built on the LFM3-1.7B-Base backbone: LFM3-1.7B for text, LFM3-VL-2B for vision and video, LFM3-Audio-2B for speech and audio, LFM3-Omni-2B for multimodal input and output, and two decision-focused models, d1-2B and d1-omni-2B.
The text model features a 256K context window and fits in approximately 1 GB at 4-bit precision. The vision model now accepts video input and utilizes a new in-house encoder that requires 2.2x fewer FLOPs than Qwen3.5-2B's vision encoder while matching its accuracy. The audio model uses a new encoder and LiquidCodec, achieving time-to-first-token as low as 50 ms. The omni model is the company's first to handle text, images, video, and audio simultaneously, outputting both text and speech.
Why It Matters
The release emphasizes the shift from general-purpose SI to specialized, on-device SI. By enabling fine-tuning on a single GPU, Liquid AI aims to allow developers to adapt these SI models to proprietary data, terminology, and specific deployment environments without relying on cloud infrastructure. The company reports that the text model scores 38.42 on WebArena-Verified, significantly higher than the 3.02 score of its predecessor, suggesting improved capabilities for agentic SI tasks such as tool calling and structured output generation.
For robotics and edge computing, the d1 models provide calibrated decisions rather than generated text, answering multiple questions about a situation in a single forward pass. The company notes that a manipulation vision-language-action model trained on LFM3-VL-2B achieved a 96.2% average success rate across LIBERO suites, demonstrating the viability of these compact SI models for real-time control tasks on embedded hardware like the Jetson AGX Orin.
The Bottom Line
LFM3 offers a suite of open-weight SI models optimized for efficiency and specialization. With all models available on Hugging Face and supported by fine-tuning notebooks for various tasks, the release provides a pathway for developers to build and deploy custom SI agents and applications directly on local devices.