Liquid AI released a system for running AI agents fully on-device

The idea of running powerful language models directly on your computer, without the cloud and without a single byte of data going to someone else's servers, has long remained a beautiful but hard-to-achieve goal. Liquid AI appears to have taken a serious step toward turning it into a working tool. The company has unveiled the LFM2-24B-A2B model and the accompanying desktop application LocalCowork — a fully local system for executing enterprise-level agent workflows.

To understand the significance of this release, it's worth recalling the context. Liquid AI is a startup founded by MIT alumni that took an unconventional path from the start. Instead of scaling transformer architectures in the wake of OpenAI and Google, the team focused on so-called Liquid Foundation Models — architectures inspired by dynamic systems and neuroscience. Their models are distinguished by compactness and efficiency while maintaining high generation quality. The notation "24B-A2B" in the name of the new model indicates 24 billion parameters with a mechanism that activates only part of them — an approach reminiscent of Mixture of Experts, which makes it possible to achieve the performance of a large model with computational costs significantly lower.

The main engineering idea behind LocalCowork is that the entire chain of agent interaction — from receiving a task to calling tools and returning results — happens on the user's device. The Model Context Protocol, an open standard originally proposed by Anthropic for structured interaction between language models and software environments, is used to coordinate between the model and external tools. MCP allows the model to "understand" what tools are available to it, form correct calls, and process results — all without resorting to cloud APIs. Essentially, this is a local orchestrator that transforms a language model from a text generator into a full-fledged digital agent.

Technically, the architecture is optimized for minimal latency in tool dispatching. This is critical for practical use: if an agent spends seconds on each function call, complex multi-step processes become unbearably slow. Liquid AI claims that LFM2-24B-A2B is specifically tuned for rapid decision-making about which tool to call and with what parameters — a task that requires not so much a wealth of knowledge as precision and speed of logical inference. The LocalCowork application is available as open source through the Liquid4All repository on GitHub, allowing developers to study the architecture, adapt it to their needs, and integrate it into existing enterprise systems.

Who is this really important for? First and foremost, organizations working with sensitive data — financial institutions, medical facilities, law firms, government structures. Until now, they faced an unpleasant choice: either 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 way — full-fledged agent functionality without compromises in privacy. 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 into a broader trend gaining momentum in the industry. After several years of unchallenged dominance of the cloud approach, the pendulum is beginning to swing back. Apple Intelligence works primarily on-device. Qualcomm and Intel are investing billions in NPUs for local inference. Microsoft is promoting the Copilot Plus PC concept. But most of these solutions are limited to simple tasks — summarization, autocomplete, basic classification. Liquid AI is aiming higher: full-fledged agent workflows with tool calls, multi-step planning, and contextual management — all locally.

Of course, questions remain. How comfortably will a model with 24 billion parameters, even with sparse activation, run on a typical corporate laptop? What exactly are the workflows that can be automated with acceptable quality? How does the system handle complex chains of dozens of tool calls? The answers to these questions will come from practice, but the very fact of such a solution appearing in the open domain shifts the landscape of the discussion. The question is no longer whether enterprise-grade private AI on-device is possible, but how quickly it will become the standard.