Liquid AI releases a system for running AI agents entirely on-device

The idea of running powerful language models on your own computer, with no cloud and not a single byte of data leaving for someone else's servers, has long been appealing but hard to achieve. Liquid AI appears to have taken a serious step toward making it a working tool. The company has unveiled the LFM2-24B-A2B model and LocalCowork, a companion desktop app: a fully local system for running enterprise-grade agentic workflows.

To grasp why this release matters, it helps to recall the context. Liquid AI is a startup founded by MIT alumni that took an unconventional path from the start. Rather than scaling transformer architectures in the footsteps of OpenAI and Google, the team focused on so-called Liquid Foundation Models, architectures inspired by dynamical systems and neuroscience. Its models are compact and efficient while keeping generation quality high. The "24B-A2B" in the new model's name points to 24 billion parameters with only a fraction of them active at a time. The approach resembles Mixture of Experts and delivers the performance of a large model at a much lower computational cost.

The core engineering idea behind LocalCowork is that the entire agent loop, from receiving a task to calling tools and returning the result, happens on the user's device. Coordination between the model and external tools relies on the Model Context Protocol, an open standard originally proposed by Anthropic for structured interaction between language models and software environments. MCP lets the model "understand" which tools are available to it, form correct calls and process the results, all without touching cloud APIs. In effect, it is a local orchestrator that turns a language model from a text generator into a full digital agent.

Technically, the architecture is optimized for minimal latency when dispatching tools. That matters for practical use: if an agent spends seconds on every function call, complex multistep processes become unbearably slow. Liquid AI says LFM2-24B-A2B is specifically tuned to decide quickly which tool to call and with which parameters, a task that demands less breadth of knowledge than precision and speed of reasoning. LocalCowork is openly available through the Liquid4All repository on GitHub, so developers can study the architecture, adapt it to their needs and integrate it into existing enterprise systems.

Who does this really matter for? Mainly organizations that handle sensitive data: financial institutions, healthcare providers, law firms, government bodies. Until now they faced an unpleasant choice: use powerful cloud models and accept that confidential data leaves the organization's perimeter, or make do with primitive local solutions. LocalCowork offers a third path, full agentic functionality with no privacy compromise. The absence of API calls means not only data protection but also independence from external services: the system works even without an internet connection.

This release fits a broader trend gaining strength in the industry. After several years of unchallenged cloud dominance, the pendulum is starting to swing back. Apple Intelligence runs mostly on-device. Qualcomm and Intel are investing billions in NPUs for local inference. Microsoft is pushing the Copilot Plus PC concept. But most of these efforts are limited to simple tasks such as summarization, autocomplete and basic classification. Liquid AI is aiming higher: full agentic workflows with tool calling, multistep planning and context management, all locally.

Questions remain, of course. How comfortably will a model with 24 billion parameters, even with sparse activation, run on a typical corporate laptop? Which workflows can actually be automated at acceptable quality? How does the system cope with complex chains of dozens of tool calls? Practice will answer these questions, but the very fact that such a solution is openly available changes the terms of the debate. The question is no longer whether private, enterprise-grade AI on a device is possible, but how quickly it will become the standard.