Run a local LLM in Unreal Engine with Ollama or LM Studio

Point Kismet’s editor agent at a model running on your own machine through Ollama or LM Studio. No API key, and nothing leaves your computer.

5 min read

Why run the model locally

A local model costs nothing per request and keeps every prompt, asset name and line of source on your machine. That matters for projects under NDA, for working offline, and for trying an agent without creating a provider account.

Kismet treats Ollama and LM Studio as providers like any other. Its agent loop, editor tools and compile checks are the same. Only the model answering changes.

Set expectations first

An editor agent is a demanding workload. Every request carries a long system prompt and the tool descriptions, and a task takes many tool calls in a row. Small models often lose track of multi-step tool use: in our experience this is common under roughly 30B parameters.

  • Pick the largest tool-calling model your hardware runs at a usable speed.
  • Give the model a context window large enough for the prompt, the tools and the conversation. Your runtime’s default may be too small.
  • Expect good results on questions and small, well-described edits before you expect them on work that spans several assets.

"Open-weight" means the weights can be downloaded, not that the license is permissive. Read a model’s license before using it commercially.

Set up Ollama

  1. 1.

    Start the Ollama server

    Kismet expects it at http://localhost:11434, Ollama’s default.

    ollama serve
  2. 2.

    Download a model in Ollama

    Note its name exactly as Ollama lists it.

  3. 3.

    Select it in Kismet

    Open the Kismet panel in Unreal Editor and type the provider command with the model name.

    /provider ollama qwen3.8:27b
  4. 4.

    Send a message

    No API key is needed. If the reply never comes, check that the server is running and that the model name matches.

Set up LM Studio

  1. 1.

    Load a model and start LM Studio’s local server

    Kismet expects it at http://localhost:1234, LM Studio’s default.

  2. 2.

    Select it in Kismet

    Use the model name exactly as LM Studio shows it.

    /provider lmstudio <model-name>

Another port, another machine, another server

The general form of the command is /provider <id> [model] [base-url]. Pass a base URL to reach a runtime on a different port or on another machine on your network, such as a workstation with a larger GPU:

/provider ollama qwen3.8:27b http://192.168.1.20:11434

You can also set OLLAMA_BASE_URL or LMSTUDIO_BASE_URL before starting the editor. For any other server that speaks an Anthropic- or OpenAI-compatible API, use the custom-anthropic or custom-openai provider with its base URL. Type /provider on its own to list the providers and whether each has a key.

Mixing local and hosted models

You are not locked into one choice. A common setup is a local model for questions about the project and quick edits, and a hosted model for larger changes. Hosted providers have sections in the model menu. Local runtimes are selected with /provider.

DeepSeek, GLM, Kimi and MiniMax appear in the model menu as hosted APIs. Kismet calls each vendor’s service for those, so they need that vendor’s key. To run an open-weight model on your own hardware, load it in Ollama or LM Studio as described above.

In the docs

Try it in your own project.

Kismet is free and open source. Install the plugin, connect a model, and ask for one small change.

Related guides