LM Studio guide: download a model and start chatting

Use the LM Studio search screen to choose a small multilingual model, checking its file size and the application's fit information. Memory use varies with model size, quantization², and format. After downloading, load the file from the model selector in the chat screen before sending a prompt³.

Request a short sentence correction to check the response. Then ask about a known item, such as an address in an installation guide, and compare with the source. Increasing the input from short chat to document questions lets you check responses, document processing, and memory in sequence.

LM Studio supports llama.cpp-based GGUF⁴ and MLX⁵ execution on Apple Silicon Mac. Two distributions with the same model name are not interchangeable files; choose for the OS and actual runtime⁶, and record model revision, quantization, context, and engine. Renaming an extension does not convert the format.

Current requirements specify Apple Silicon and macOS 14 or later on Mac; Intel Mac is unsupported. On Windows x64, check AVX2 support. Application installation requirements differ from the memory needed for a model and long documents. If loading fails, check model size, available memory, app/engine version, and file format.

The CLI⁷ option lms load --estimate-only estimates memory without loading the model, but it does not guarantee usage in other applications or runtime behavior. First load the model with a short context and confirm basic chat, then add the required document input. Estimation, actual loading, and document processing are separate stages.

The example starts a server with the alias local-guide and a 4,096-token context. Send a short prompt first, then increase context only when the actual task requires it. This helps distinguish model-loading failure from document text missing in the API request.

Running lms load opens a selection of downloaded models. Choose the file and wait for loading to finish. You can inspect the files first with lms ls. The command assigns the API alias local-guide to the selected model; use that alias in the next request.

Start the server and send the request below. local-guide is the API alias from the previous step, not a filename. Both curl and the client should use 127.0.0.1:1234 and a supported API path. A returned sentence confirms model selection and server connectivity, but does not mean the guide attached in the interface was sent to the API.

If authentication is enabled, include its token⁸ in the request header. Use 127.0.0.1 for testing on the same computer; do not expose the server to the network or enable CORS as a diagnostic shortcut. Missing document content is not fixed by changing the address or CORS.

Send the same short text question in the interface and through the API. If both behave similarly but document questions differ, check whether the API input omitted document content. If even short questions differ, inspect the model alias, chat template, and generation settings.

If API answers about the guide are wrong, check whether the application actually included the document in its request. Short material can be included directly; longer material may need retrieval of relevant passages. Processing triggered by attaching a file in the interface is not automatically carried over to API calls.

To investigate, include a sentence you already know from the guide directly in the API request and ask about it. If the model answers from that sentence, the model and API are processing text. If results differ only when you attach the file in the UI, check how the app retrieves document text and adds it to the API request.

If the answer format is still wrong with the text included directly, compare the model file, chat template, and generation settings. This test diagnoses the path used in this case; it does not measure general document-retrieval accuracy.

Test tool use separately from ordinary chat. Receiving a request, generating a tool-call format, executing the call in the application, and returning its result to the model are separate stages. After plain chat, add one needed tool, then connect the document task to isolate the failing boundary.

If the API server must keep running after closing the application interface, consider LM Studio's separate llmster daemon. Using an API alone does not require switching products. A daemon needs its own startup, update, and restart-recovery plan, so a GUI may be simpler for short tasks that you supervise directly.

Before comparing acceleration or file support in another engine, record the model file, runtime, context, and request that produced the current answer. If including document content in the API solves the task, you can keep LM Studio. If retrieval/insertion is missing, improve the client; consider changing runtimes only when the current engine lacks a needed format or acceleration.