Build specialized multimodal models with LFM3

A Unified Backbone for the Edge

LiquidAI has released its LFM3 family, a new generation of multimodal foundation models designed specifically for specialization and on-device deployment. The series includes six models in the 2-billion parameter range—spanning text, vision, audio, and omni-modal capabilities—all built from a single shared backbone. This approach enables developers to select a modality-specific model, fine-tune it on custom data using a single GPU, and deploy it directly on edge hardware, representing a significant push towards accessible and efficient local AI.

Technical Architecture and Performance

The core of the new family is the LFM3-1.7B-Base model, which leverages a new tri-hybrid architecture that outperforms its larger predecessor, LFM2.5-2.6B, while using 35% fewer parameters. This efficiency is achieved through a mix of short convolutions, sliding-window attention, and global attention layers, a design refined through hardware-in-the-loop architecture search. For multimodal tasks, LiquidAI also introduced new in-house vision and audio encoders that reduce computational overhead, with the vision encoder matching the accuracy of a competitor's model with 2.2x fewer FLOPs. The text model supports a 256K context window, enabling more complex agentic tasks to run locally.

• Shared Backbone: All models branch from the LFM3-1.7B-Base, trained on 22T tokens.
• Parameter Size: Models are kept within the 2B parameter class for edge compatibility.
• Architecture: A tri-hybrid mix of gated convolutions and attention mechanisms.
• Context Length: The text model features a 256K context window for on-device agents.
• Efficient Encoders: New in-house vision and audio encoders reduce compute and latency.
• Accessibility: Every model is designed to be fine-tuned on a single consumer-grade GPU.

The release challenges the industry's focus on scaling large, general-purpose models by providing a viable framework for building decentralized and specialized intelligence. By optimizing for performance-per-parameter and providing tools like the new LiquidCodec audio tokenizer for fast speech generation, LiquidAI is equipping developers to build more private and responsive applications that are not reliant on cloud APIs. This could influence the development of next-generation robotics, voice assistants, and visual understanding tools that operate entirely at the edge.

LiquidAI's LFM3 family represents a calculated bet that the future of applied AI lies not in scaling up massive, centralized models, but in scaling out highly efficient, specialized models that can be customized and deployed directly at the edge. This strategy prioritizes developer accessibility, application-specific performance, and on-device privacy over chasing parameter count.