LM Studio

LM Studio is a desktop app for finding, downloading and running models with a graphical interface. It includes an OpenAI-compatible local server that twinny can use for all three jobs.

Set up

1. Install LM Studio and download a model from its Discover tab. For chat pick an instruct model (Qwen2.5-Coder-7B-Instruct); for completion a base or code model (Qwen2.5-Coder-1.5B); for embeddings an embedding model (nomic-embed-text).
2. Open the Developer tab and start the server. The default port is 1234.
3. Load the model(s) you want to serve. LM Studio can load several at once when memory allows, and can load on demand if just-in-time loading is on.

In twinny

Add provider → LM Studio. twinny asks the server for its model list and fills in the first match for the job. Defaults:

Job	Path
Chat	/v1
Autocomplete	/v1/completions
Embeddings	/v1/embeddings

The model name is the identifier LM Studio’s server reports, which you can see in the Developer tab; it is often the repository name in lower case, e.g. qwen2.5-coder-7b-instruct.

Notes

• Autocomplete uses the OpenAI-style /v1/completions route with a FIM prompt rendered for the model. Automatic template detection works from the model name; if the name does not contain the family (say a renamed GGUF), pick the template by hand.
• To use LM Studio from another machine, turn on Serve on Local Network in the server settings and point the provider at the machine’s address. Or run Ollama on that machine and use Devices.
• The server’s context length is set per model in LM Studio’s model settings. Raise it for review and long chats.
• If requests fail with a model-not-found error, the model is not loaded. Load it in LM Studio or enable just-in-time loading.