Liquid AI Enables Privacy-First Local Agent Workflows

Liquid AI, the company behind the new framework, spun out of MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL). The startup, founded in 2023, has raised a total of $297 million and reached a valuation of over $2 billion after its $250 million Series A round in December 2024, which was led by AMD Ventures. The LFM2-24B-A2B model's efficiency stems from its Sparse Mixture-of-Experts (MoE) architecture. While it contains 24 billion total parameters, only about 2 billion are active for any given token, allowing the model to fit into as little as 14.5 GB of RAM when quantized. On-device performance tests on an Apple M4 Max showed the model achieving tool-selection responses in approximately 385 milliseconds. This low latency is ideal for interactive, human-in-the-loop applications rather than fully autonomous, multi-step workflows where its end-to-end success rate is currently around 26%. The system uses the Model Context Protocol (MCP), an open standard for enabling AI models to interact with external tools and data sources. Think of it as a universal connector, or a "USB-C port for AI," designed to standardize communication between an AI and local applications like your calendar or file system. Liquid AI's core technology is intentionally different from the dominant Transformer architecture. The founders are developing what they call Liquid Foundation Models (LFMs), designed with a hardware-in-the-loop approach to optimize for speed and efficiency on edge devices like laptops and phones, not just in the cloud. The LocalCowork desktop application is open-source and comes with pre-built tools for file operations and security scans, such as identifying leaked API keys or personally identifiable information (PII) on the host machine. The model supports inference through widely-used frameworks like llama.cpp and vLLM, and its code is available on GitHub.