Liquid AI Releases LFM3-1.7B Model with Tri-Hybrid Architecture
Liquid AI has released LFM3-1.7B, a 1.73-billion parameter text model featuring a tri-hybrid architecture combining gated short convolutions and attention mechanisms. The model supports a massive 256K-token context window and is specifically optimized for local agents and edge devices. This release marks a significant step forward for on-device AI, enabling powerful agentic workflows and large context processing directly on consumer hardware like smartphones and laptops. At 4-bit precision, it consumes about 1 GB of memory, making advanced local AI much more accessible. The model features 24 layers split into 13 gated short convolutions, 5 sliding-window attention, and 6 global attention layers, alongside a 128,000-token vocabulary. While highly efficient for tool calling and structured outputs, the developers advise against using it for complex multi-step reasoning or agentic coding due to its compact size.
## BACKGROUND
Liquid Foundation Models (LFMs) developed by Liquid AI are designed for fast inference and on-device deployment by integrating structured architectural components like convolutions with attention mechanisms. The GGUF format is widely used to store and run large language models efficiently on consumer hardware such as local PCs and edge devices.