Liquid AI unveils LFM3 family of customizable AI models for local deployment
Summary
Liquid AI has launched its new family of six customizable AI models known as LFM3, which encompasses capabilities for text, images, video, voice, and structured decision-making, all designed to operate on a single GPU. Notably, the 1.9 billion parameter vision model within this suite performed comparably to Gemma 4's 12 billion parameter model, highlighting the efficiency of LFM3’s architecture. This launch allows developers to tailor these models to their specific needs and operate them locally, circumventing ongoing API costs. The models support multimodal integration, making them suitable for creating more natural and context-aware AI assistants that can handle various input types effectively.
Analysis
LFM3: LFM3 is Liquid AI's new generation of foundation models spanning text, vision, audio, omni, and decision modalities. It incorporates a tri-hybrid architecture combining gated short convolutions, sliding-window attention, and global attention for efficient long-context processing. The family is designed for fine-tuning and on-device deployment across multiple processor types. d1-2B: d1-2B is a decision model in the LFM3 family that produces structured outputs with calibrated probabilities in a single pass for actionable application results. @liquidai: @liquidai is the official X account of Liquid AI used to announce new model releases, architectural details, and deployment capabilities. The account shared the LFM3 launch thread highlighting the tri-hybrid architecture and multimodal support. LFM3-1.9B: LFM3-1.9B is a text-focused model in the LFM3 lineup with a 256K-token context window suitable for agentic tasks and document processing. It supports tool use across browsers and business applications. Liquid AI: Liquid AI develops efficient foundation models optimized for multiple modalities including text, vision, audio, and decision-making. The company focuses on architectures that enhance performance per parameter while supporting local inference across diverse hardware. It recently introduced the LFM3 family to enable developers to customize and deploy specialized AI assistants on their own hardware. LFM3-VL-2B: LFM3-VL-2B is a vision-language model within the LFM3 family that handles images, video, charts, and spatial reasoning. It supports video input for enhanced scene understanding and object localization. d1-Omni-2B: d1-Omni-2B is an omni variant of the d1 decision model that adds audio input support while maintaining structured, probability-calibrated outputs. LFM3-Omni-2B: LFM3-Omni-2B is an omni model in the LFM3 family that integrates vision, audio, and language capabilities for multimodal interactions where the assistant can respond to combined inputs. LFM3-Audio-2B: LFM3-Audio-2B is an audio model in the LFM3 suite enabling voice conversations, tool calls, and reasoning across 18 languages. Hardware Efficiency: Architectures optimized for diverse processors including GPUs, CPUs, and NPUs support deployment on controlled enterprise hardware. On-Device Deployment: Developers can adapt foundation models to proprietary data and tools for local inference without ongoing per-token API costs. Multimodal Integration: Models combining vision, audio, and language inputs enable more natural and context-aware AI assistants across languages and media types.
Categories
aimachine_learningtech