On September 15, LM-Kit One becomes generally available: signed packages for Windows, Linux, and macOS, free to build and evaluate, with nothing to activate. The LM-Kit One page covers what ships, and the documentation is already live. Here is what ships, why we built it, and what we learned running it ourselves for the past year.
Serving a model is the easy part #
Local inference has become remarkably accessible, and that is good news. But serving tokens is only one piece of putting AI into a real application. Teams still have to assemble retrieval, document processing, search, agents, tools, identity, policy, audit, and administration around it, and then live with the seams between a dozen separately versioned parts.
LM-Kit One exists to collapse that stack. It is a private AI application server: one server carrying local inference, governed agents, RAG with cited answers, multi-tenant search, document intelligence, and the operational layer around them, developed and released as one product. Compatible clients connect through the interfaces they already speak: OpenAI, Anthropic, Ollama, and MCP compatible endpoints, plus a native REST API for what those dialects cannot express.
We are deliberate about the edges of that promise. We publish the endpoint-by-endpoint API coverage, including what we do not support. Storage is composable rather than prescribed: run entirely locally with no required external service, or connect LM-Kit One to the relational databases and vector stores your operators already manage and back up. Security ships in the server itself, identities, per-key policies, and audit, and external access stays explicit and controlled. Grounded-answer workflows are designed to abstain when the evidence falls below the configured threshold. That is sovereignty in practice: your infrastructure, your jurisdiction, your rules.
What ships on the 15th #
This is a release, not a waitlist. Every package carries the full server, and no LM-Kit account, activation key, or runtime license check is required to get started.
On the download page, September 15
- Five downloadable packages: signed installers and portable archives for Windows x64 and Arm64, self-contained archives for Linux x64 and Arm64, and a notarized package for Apple Silicon.
- CUDA, Vulkan, and Metal acceleration, with a CPU AVX/AVX2 fallback.
- Runs as a Windows service, a desktop session, or a self-contained Linux server. Start on one machine and scale out horizontally, any node serving any request, without changing the application stack.
- Free in production, commercial use included, below the published thresholds ; special terms for education and research; evaluation free at any size, with no time limit.
We ran it on ourselves before asking you to #
LM-Kit is part of the Calico IIM Group, and for more than a year LM-Kit One has been the AI stack behind every software product the group operates. Two things from that year say more than any feature list.
-
We retired our search clusters. The group's products ran on OpenSearch; they now run on LM-Kit One's built-in search service, inside our own perimeter. Across our production workload, the migration divided the cost of search by 10 while improving both accuracy and latency. The same backbone now powers keyword search, RAG with cited answers, and the retrieval our agents rely on, and one whole external system, with its licenses and upkeep, left the architecture.
-
Agents took the repetitive half. More than half of the repetitive operational workflows we have identified for automation now run through agents served by LM-Kit One, under explicit permission policies, with every action audited. As the server's intelligence advances release after release, we expect LM-Kit One to be carrying 90% of those workflows next year. The value has been straightforward: less time on repetitive operations, more time for the work that requires judgment.
The release you download on the 15th is the stack we run ourselves. Same packages, same engine, same defaults.
Three years to get here #
LM-Kit started three years ago inside the Calico IIM Group, four merged companies whose teams have served the enterprise content management market for more than 40 years. The bet was simple to state and hard to build: enterprise AI would increasingly run beside the data it works on, under the laws and the controls of the organization that owns it, rather than pushing every prompt, document, and workflow across a cloud boundary. We spent those years rebuilding the layers that make that practical: inference, document processing, retrieval, search, agents, and the operational controls around them. More than 150 releases and 23,000 commits later, that stack became LM-Kit One.
The stack moves as one: each release carries every layer together, so when the runtime, a document engine, or search improves, everything running on the server inherits it in the same release. That cadence is deliberate. Intelligent information management is our trade, and in that trade a stack is only as strong as its weakest layer, so every layer gets pushed, measured, and shipped as part of one engine.
3 years of development and research
150 releases since 2024
+ 360k downloads across NuGet packages
+ 1 year in production at Calico IIM
+ 40 years of combined ECM experience
+ It also carries our next act. On this same server we are building Burofolk, an AI-native workspace where people and agents work side by side across documents, projects, and everyday operations, and we will have more to share about its release soon. Everything we learn building it lands back in the server you can download today.
LM-Kit One is available #
The signed packages are live on the download page, and the launch post covers what ships and what it costs, which for evaluation is nothing. The documentation and the complete REST API reference are open.
LM-Kit One is available now
Five downloadable packages, free to build and evaluate, nothing to activate.
Already operating a model server, a vector database, a retrieval pipeline, and a document stack separately? We would like to hear what you are running.