LM-Kit One · Coming from Ollama

Keep the client. Change the host.

Anything written against the Ollama API, including the stock CLI, works when the host points at LM-Kit One. Then the application stack starts.

Full surface, lifecycle included Your GGUF models import Free to build and evaluate

One environment variable, or one host argument.

The dialect is the door, not the product: the verbs you already use land on a bigger server.

The unmodified ollama CLI, aimed at your server.

terminal

# Point the stock client at LM-Kit One
export OLLAMA_HOST=http://your-server:5189

ollama pull qwen3.5:9b
ollama run qwen3.5:9b "Summarize this contract clause."

The Ollama Python library, with only the host changed.

app.py

from ollama import Client

client = Client(host="http://your-server:5189")
reply = client.chat(
    model="qwen3.5:9b",
    messages=[{"role": "user", "content": "Summarize this contract clause."}],
)
print(reply["message"]["content"])

The management verbs work too, which is the part compatibility shims usually skip.

terminal

ollama list          # tags from this server's store
ollama ps            # what is loaded right now
ollama create mine -f Modelfile   # import your own GGUF
ollama push mine     # multi-gigabyte uploads are a tested path

Chat, generate, embeddings, tags, show, ps, version, pull, push, create, copy, delete and blob upload are all served; the coverage matrix is the authoritative list.

The moment a runner stops being enough.

It worked on your machine. Now five applications and a team need it, and the model runner was never the hard part.

Access

Callers become identities

Tokens hashed at rest, per-caller isolation, request history and audit, instead of an open port everyone shares.

Documents

The workload beyond chat

OCR, conversion, structured extraction, search collections and answers with page citations, on the same server.

Agents

Behaviour defined server-side

Named agents with skills, governed tools and memory, adopted by any client with one field instead of rebuilt per app.

Dialects

Three doors, one engine

The same models also answer OpenAI and Anthropic clients and MCP assistants, so every tool in the house lands on one governed server.

Scale

Past one machine

Horizontal scaling for inference and sustained document workloads when the experiment becomes infrastructure.

Operations

A console, not a terminal

Live telemetry, request attribution, model fit against hardware, alerts and logs, for the person who has to keep it up.

Worth knowing before you switch.

Two differences, stated here rather than discovered later.

Catalog

Names map to this server's store

Model names resolve against the LM-Kit catalog and your imports, not Ollama's registry. Pull what you need, or bring your own GGUF through create and push.

Purpose

It is a server, not a tray app

LM-Kit One is built to be operated: identities, policies and an admin console come with it. For personal single-machine use, a runner may stay the simpler tool.

LM-Kit One

Repoint one client and see for yourself.

Download LM-Kit One Follow the quickstart