LM-Kit One is here: private AI, under your control.

LM-Kit One brings models, agents, search, document intelligence, and the controls needed to operate them into one product, running on infrastructure you control. The packages for Windows, Linux, and macOS are on the download page now.

We built it around a straightforward requirement: organizations should be able to put their knowledge to work with AI without handing over control of it. Getting started should be accessible. Moving into production should not mean rebuilding the application around a different stack.

With LM-Kit One, you can extract structured data from documents, build assistants that answer from internal knowledge with source references, and give agents tools to complete work under explicit permissions. These capabilities are developed and released together, rather than left to each team to assemble and maintain separately.

The problem is rarely the model

Running a model is only the beginning. Putting it to work on private data also requires document processing, retrieval, tools, access controls, and monitoring. Building and maintaining those connections can become a project of its own, before the team has delivered the application it set out to build. Most of that work is the same from one organization to the next. LM-Kit One does those jobs as one server, built, versioned, and improved together.

It is for every developer and team that needs private AI, in two situations that often meet in the same organization:

• Putting AI to work inside the organization. A document process that needs structured data out of scans and PDFs, an assistant that answers from company knowledge with sources, agents on repetitive operational workflows: all of it running on servers the organization controls, for people who cannot send that information anywhere else.
• Building private AI into the software you ship. Software vendors, integrators, and in-house product teams get one backend for extraction, search, agents, and chat that installs beside their application at the customer, reached through the native REST API or the OpenAI-, Anthropic-, and Ollama-compatible interfaces and MCP, so compatible applications and clients connect through familiar interfaces, with supported endpoints documented in the compatibility matrix.

In both cases, the IT and platform team that runs it gets what it needs to say yes: identity, access policies, audit, and an explicit boundary on what may leave the network.

Sovereignty is an operational choice

Sovereignty means more than choosing where a server is hosted. It means deciding which models run, who can access information, what an agent is allowed to do, and whether anything may leave your environment.

LM-Kit One supports fully local operation, including reasoning, document processing, indexing, retrieval, and fine-tuning. Once the models are provisioned, it runs on an isolated network without an external AI service. Connecting external tools or assistants remains your decision, with an important distinction: an external assistant receives the results of the tools you allow it to call. Fully private operation does not require those connections.

The server does not send us your documents, prompts, results, or usage telemetry. Operational logs and metrics remain available to you on your server. Telemetry export is disabled unless an operator enables it. Access policies, administrative roles, and audit records give your team the controls to manage the deployment itself. The Trust page states each of these commitments in full.

The point is not simply to keep information confidential. It is to make AI something your organization can operate, inspect, and govern.

Start on one machine. Expand when the work demands it.

Private AI should not stop at a desktop experiment.

LM-Kit One begins as a single-machine deployment and grows into multiple nodes behind a load balancer, with shared databases and storage. Kubernetes is an option, not a requirement: the deployment model also accommodates virtual machines and Windows services. Scaling adds processing capacity while the application backend stays the same.

You choose the models and hardware, measure them against your workload, and expand where needed. A good starting point is one document workflow or one internal assistant. The architecture supports what comes next without forcing you to move the data somewhere else.

Built by people who manage information for a living

LM-Kit is part of the Calico IIM group, whose companies bring more than 40 years of enterprise content management experience and serve over 600 customers across public institutions, education, healthcare, and businesses working with information they cannot afford to mishandle. In that environment, a plausible answer is not enough. Information needs to remain accessible to the right people, traceable to its source, and usable inside an existing business process. Those requirements matter as much as the model.

LM-Kit One has run inside the group for more than a year, powering its software and operational workflows. The release you download today is that same stack. We wrote about what that year taught us when we announced the date.

A sustained investment in private AI

We are investing heavily in the engineering behind private AI: more efficient inference, better document understanding, more accurate retrieval, and more reliable agents. Our commitment is to keep improving what organizations can achieve on infrastructure they control, without making every team rebuild its stack as the technology evolves. Because these layers ship together, each improvement reaches the applications built on the server in one coordinated release, the cadence behind more than 150 releases of the LM-Kit engine since 2024.

Our objective is to make cutting-edge AI practical for any organization, anywhere in the world, while keeping it fully private and under that organization's control. That means pursuing better results and more efficient use of hardware, while preserving the ability to choose models, control access, and decide when to adopt changes. Innovation should expand what an organization can do, not reduce its independence.

Free to evaluate. No trial gate.

Testing and development are completely free, at any organization size, with no time limit. You get the full server, without an LM-Kit account, license key, activation process, or trial countdown. You do not need to speak to us before getting started.

Production use, including commercial use, is also free for businesses that meet all three conditions:

• Gross annual revenue below US$1 million.
• 10 employees or fewer.
• Total outside capital received of US$3 million or less.

Eligibility is assessed together with parent, subsidiary, and affiliated entities. Businesses outside these limits need a Professional License for production use. For client projects, eligibility follows the client, not the contractor. Personal use, education, registered nonprofits, and open-source projects also receive free-use rights under the EULA. The models you choose carry their own licenses, which apply separately.

One boundary matters for software vendors: embedding LM-Kit One in an application that adds substantial functionality of its own is permitted under the EULA. Developer-facing products and general-purpose AI APIs require a separate agreement.

Bring a real project

Start with work you actually need to do: a document process, a knowledge collection, or an application that needs private AI. The quickstart walks through three outcomes worth proving in your first hour: a document becomes structured data, a folder becomes searchable knowledge with citations, and a client you already own points at your own server. Evaluate it with your own data, on your own infrastructure, against your own requirements.