LM Studio Not Loading Models? How to Fix “Failed to Load Model” and Search Errors

LM Studio says “Failed to load model”, the progress bar hangs forever, the model you downloaded does not show up in My Models, or the search box finds nothing at all.

These problems look different, but they come from a short list of causes: not enough memory for the settings you picked, an outdated runtime, a file in the wrong folder or only half downloaded, or a network block on Hugging Face.

This guide shows how to read the error, then fixes each cause step by step, with the exact settings, folders and lms commands. It works for Windows, macOS and Linux.

Step 1: Read the Real Error

The pop-up message is often vague. The useful detail is in the log.

1. Open the Developer tab (the terminal icon in the left bar). If you do not see it, switch the interface mode to Power User or Developer at the bottom of the window.
2. Try to load the model again and watch the log.
3. Find the first line that starts with llama_model_load or mentions an error.

You can also stream the log in a terminal with lms log stream. Once you have that line, the table below points you to the fix.

Quick Diagnosis Table

Fix 1: Not Enough VRAM or RAM

This is the most common cause. A model needs memory for its weights plus the KV cache, and the KV cache grows with context length. A model that loads at 4K context can fail at 32K.

Work through these in order, reloading after each change:

1. Lower Context Length. In the load settings panel (the gear next to the model selector, or the right sidebar when you pick a model), drop it to 4096 or 8192 to test.
2. Lower GPU Offload. Move the slider down so some layers run on the CPU. Slower, but it loads.
3. Close other GPU apps, such as games, ComfyUI or a second LM Studio model. Unload models you are not using.
4. Pick a smaller quant. A Q4_K_M file is far smaller than Q8_0. Our GGUF quantization guide explains which to choose.
5. Turn on Flash Attention if the model supports it, which cuts KV cache memory.

A quick test from the error guide on DEV Community: if the error changes when you move the sliders, it is a memory problem. If it fails the same way at every setting, it is something else.

On Apple Silicon, memory is shared with macOS and your other apps, so an out-of-memory failure shows up as a Metal buffer error. Close browsers and try a smaller context.

As a rough guide for weights alone, a 7B to 9B model at Q4 needs about 5 to 6 GB, a 27B to 32B model at Q4 about 16 to 20 GB, before any context. For model picks by card size, see our best local models by VRAM list, which applies to LM Studio too.

Fix 2: The Loading Guardrail Blocks It

LM Studio has model loading guardrails that refuse to load a model it thinks will not fit, to stop your system freezing. If you see an “insufficient system resources” message, the guardrail stopped it before it tried.

• The error notification links straight to the guardrail setting (added in v0.3.9).
• You can also find it in App Settings (Ctrl/Cmd + comma).
• Set it to a less strict level, or off, only if you know the model will fit, for example because you plan to lower the GPU offload.

Relaxing the guardrail does not create memory. If the model really is too big, it will fail later with an allocation error or make your PC very slow. Fix 1 is still the real answer.

Fix 3: Unsupported Architecture or Old Runtime

LM Studio runs GGUF models through a llama.cpp engine (and MLX on Macs). These runtimes update separately from the app. A brand-new model often needs a newer runtime.

1. Open the runtime manager with Ctrl + Shift + R (Windows and Linux) or Cmd + Shift + R (macOS).
2. Update the runtime you use. On NVIDIA that is usually the CUDA llama.cpp runtime. Vulkan works on most GPUs, including AMD and Intel.
3. Update the app itself from the menu or the website.
4. Restart LM Studio and try again.

Runtime auto-update is on by default, but it can be switched off in App Settings, so check it.

If the architecture is still unknown after updating, llama.cpp itself may not support the model yet. No LM Studio setting can fix that. Wait for support, or use a runtime that already has it. Our Ollama vs LM Studio vs llama.cpp guide explains how they relate.

If one runtime crashes on your GPU, try another. For example, switch from CUDA to Vulkan as a test.

Fix 4: Broken, Partial or Split Files

A download that stopped early, or a file that got damaged, gives errors about magic numbers, tensor counts or reading past the end of the file.

1. Compare sizes. Check the file size on disk against the size listed on the Hugging Face page.
2. Check split models. Big models come in parts named like -00001-of-00003.gguf. All parts must be in the same folder, and you load the first part.
3. Delete and download again. Do not resume a file you think is damaged.
4. Free disk space. A full disk is a common reason downloads stop part-way.
5. Check antivirus. Some security tools lock or quarantine large files while scanning them. Add an exception for the models folder if this keeps happening.

Fix 5: Model Not Showing in My Models

LM Studio only finds models in a fixed layout. From the LM Studio import docs, the structure is:

~/.lmstudio/models/
  publisher/
    model/
      model-file.gguf

On Windows, ~ is your user folder, so the path is C:\Users\YourName\.lmstudio\models\. The file must be exactly two folders below the models folder. A GGUF dropped straight into models, or three levels deep, may not appear.

Ways to fix it:

• Move the file into models/your-name/model-name/. Any folder names work, as long as there are two.
• Use the import command: lms import path/to/model.gguf, then follow the prompts. LM Studio labels it experimental.
• Check which folder LM Studio uses. You can change the models directory from the My Models tab. After an update or moving drives, it may point somewhere else.
• Check the file type. LM Studio loads GGUF everywhere and MLX on Apple Silicon. A .safetensors model meant for PyTorch will not load.

If you also use Ollama, its model files are stored in a different format and are not picked up by LM Studio automatically. See our Ollama commands cheat sheet for where Ollama keeps them.

Fix 6: Vision Model Loads but Images Fail

Vision models in GGUF form usually need two files: the main model and a projector, often named mmproj-*.gguf. If the text side works but images give an error or are ignored, the projector is missing.

• Download the matching mmproj file from the same repo.
• Put it in the same folder as the model.
• Use a projector made for that exact model. One from a different model or size will not work.

Downloading through LM Studio’s own search usually fetches both files for you.

Fix 7: Model Search Not Working

LM Studio works offline once a model is downloaded, but the docs say plainly that search will not work without an internet connection, because it queries Hugging Face.

1. Check you can open huggingface.co in a browser on the same PC.
2. Turn on the Hugging Face proxy in App Settings. LM Studio added “an option to use LM Studio’s Hugging Face proxy” in v0.3.9 for people who cannot reach Hugging Face directly.
3. Check firewalls, VPNs and work networks. Some block Hugging Face or large downloads.
4. Search by the exact repo name, such as unsloth/Qwen3.8-27B-GGUF, or paste the Hugging Face link into the search box.
5. Download manually from the Hugging Face website, then place it as in Fix 5.
6. Use the CLI: lms get followed by a keyword downloads from the terminal.

If a model you expect is missing from the results, check its file list on Hugging Face directly, because some repos only hold formats LM Studio cannot load.

Fix 8: Hardware and OS Requirements

Some load failures are simply unsupported hardware. From the LM Studio system requirements:

Intel Macs are not supported, and older x64 CPUs without AVX2 will not run it. For AMD graphics cards, see our AMD GPU for AI guide.

Useful lms Commands

The lms command-line tool ships with LM Studio and is handy for testing outside the app.

Run lms --help for the full list on your version. For agents and MCP tools inside LM Studio, see our LM Studio MCP and Agent Mode guide.

FAQ

Why does LM Studio say “Failed to load model”?

Most often the context length and GPU offload need more memory than you have. Read the log: an allocation error means memory, an architecture error means an old runtime, and a magic-number error means a bad file.

How do I update the LM Studio runtime?

Press Ctrl + Shift + R (Cmd + Shift + R on Mac) to open the runtime manager and update the llama.cpp or MLX runtime. Then update the app.

Where does LM Studio store models?

In ~/.lmstudio/models/ by default, in a publisher/model/file layout. You can change the folder from the My Models tab.

Why is my downloaded GGUF not showing up?

It must be exactly two folders below the models directory. Move it into models/publisher/model/ or use lms import.

Why does model search show nothing?

Search needs internet access to Hugging Face. Check your connection, try the Hugging Face proxy option in settings, or download the file manually.

Can LM Studio run on an Intel Mac?

No. The macOS version needs Apple Silicon and macOS 14 or newer.

Sources

• LM Studio docs: Import models (folder layout, lms import)
• LM Studio docs: System requirements
• LM Studio v0.3.9 release notes (Hugging Face proxy, guardrail quick access, runtime shortcut)
• DEV Community: LM Studio failed to load model (log lines and order of checks)