Quick diagnosis
A downloaded model and a loadable model are different things. If LM Studio cannot load a file, first check the download type, available memory, and selected inference engine. A larger download or repeated reinstall does not fix an incompatible model architecture or an allocation failure.
This guide suits a model that appears in My Models but fails before a short chat begins. If the app itself cannot start, begin with its system requirements and the correct installer for your operating system and processor.
First three checks
- Copy the exact error, model publisher, file name, compression option, and selected engine. Keep personal prompts out of screenshots and logs.
- Unload other models, close memory-heavy apps, and try one short new chat with a modest context such as 4096 tokens.
- Confirm that the download belongs to an architecture supported by the selected engine. GGUF and MLX downloads are not interchangeable.
Success is a model that finishes loading and answers a short prompt while the computer remains responsive. A completed progress bar alone does not establish that a document workflow or a long conversation will fit.
Check the file and engine together
LM Studio runs supported GGUF models through llama.cpp on Mac, Windows, and Linux. Its MLX engine supports Apple Silicon. A Windows computer cannot load an MLX package simply by selecting a different compression label. Use the GGUF vs MLX comparison when the download options are confusing.
A .gguf extension also does not mean that every architecture is supported by your installed engine version. Check the publisher's model card and required runtime support. Open LM Studio's runtime manager with Ctrl+Shift+R on Windows or Linux, or Cmd+Shift+R on Mac, and check the installed engine. Update through the app's supported route when the error identifies missing architecture support.
For an imported model, preserve the expected publisher/model folder structure. A split download needs all its parts; a vision model may need companion files and matching engine support. An incomplete or unrelated file cannot be repaired by increasing context or GPU offload.
Check memory before raising settings
Model weights, conversation memory, runtime buffers, and the desktop all consume memory. A 5GB file is not a promise that a 6GB GPU can hold the full workload. Enter the exact file size, hardware, and intended context in the memory calculator.
Try a smaller supported model file first. If that works, return to the larger model with shorter context and an offload setting appropriate to your hardware. More GPU offload can help speed when the GPU has room; it can make loading fail when its dedicated memory is already full. Partial CPU offload also needs enough system RAM.
If you use the optional lms CLI, list the downloaded models and preview the memory requirement without loading one:
lms ls
lms load "MODEL_KEY" --context-length 4096 --estimate-only
Replace MODEL_KEY with the key shown by your own list. This is an estimate for the selected load settings, not a measured speed result or a guarantee that every other app will leave enough memory free. The desktop app can be used without adopting the CLI.
Match the symptom to the next action
| Symptom | Useful next action |
|---|---|
| Out of memory or allocation failure | Unload other models, reduce context, and try a smaller file |
| Unsupported architecture or tensor format | Check the model card and update or select a compatible engine |
| Model missing after changing drives | Check My Models' directory, permissions, and available drive |
| Interrupted download or missing split file | Complete the download before changing load settings |
| Small model works but large one fails | Recalculate the full memory requirement; keep the working model available |
| Every compatible small model fails | Check app requirements, engine availability, and the exact diagnostic error |
For a GPU detection error, use the GPU diagnostic. For an unavailable drive, use moving model storage. A model that loads but gives bad answers needs the answer-quality checks, rather than another installation.
Offline and document limitations
LM Studio's core local chat can work offline after the model and engine are available. Downloading a new model or required engine needs connectivity. Prepare those before disconnecting; an unavailable download is different from a failed local load.
Once a short chat works, add one small document and check a known fact. Larger context, document retrieval, vision, or a second loaded model can introduce another memory or compatibility problem. Restore those features one at a time so the failing condition remains identifiable.