Run a GGUF model in your browser with llama.cpp

If you want to try a local LLM using GGUF, Naidan can run a compatible chat model in your browser. Choose from the models below. To use a different GGUF, open Settings to import a file from your device or download one from Hugging Face.

Choose a model to try in your browser

Choose a model to open its chat. If it has not been downloaded, select Start with this model there; once it is ready, enter your message. Local-only cannot fetch an unprepared model from Hugging Face: import a GGUF you obtained separately, or download the model in a Hosted edition that permits the required model file hosts.

To make your own model link, set the app URL’s llama-cpp-browser-model parameter to an hf.co GGUF model reference such as hf.co/owner/repository:Q4_K_M.

The llama.cpp browser option appears in both Hosted and Standalone. Inference settings checks what this browser and device support. Model compatibility and speed depend on the device and model.

1. Open the browser model manager

To import your own GGUF or download another model, open /#/settings/llama-cpp-browser to reach the model manager.

If first-run setup appears, close it with the × at the top right. The linked llama.cpp browser settings tab stays open for the steps below.

The link above opens the llama.cpp browser tab directly. You can also open it from Settings in the sidebar. Check Inference settings for browser support. Automatic chooses among available profiles; importing a GGUF cannot make an unavailable profile run.

Settings opens the llama.cpp browser manager and shows runtime support.

2. Import a GGUF you already have

For a local model, select Choose GGUF files or Choose model folder in the GGUF model files area. You can also drop files or folders there. After import, the file appears under Imported models. If you are downloading from Hugging Face instead, continue to the next step.

The chosen GGUF appears in Imported models after the import finishes.

3. Or download a GGUF from Hugging Face

On Local-only, import a GGUF obtained separately as in the previous step. In an edition that permits connections to the model file hosts (Curated permits huggingface.co and *.hf.co), enter an owner/repository name and choose Check model. Review the available quantization and, if needed, its model variant. Under Download to, choose Browser storage (OPFS) or a linked device folder that your browser supports, then select Download. The manager adds the finished files to Imported models. For image chat, a matching mmproj GGUF—the model’s companion file for image input—is also needed.

Check the repository, choose files and destination, then download before using the model.

4. Choose the default model for chats

In Imported models, select Set as default beside the model and confirm. This saves both llama.cpp browser as the global endpoint and the selected model. Existing and new chats that inherit those values use the choice. Importing or downloading alone does not save it, and an individual chat can select another model.

The confirmation changes the global endpoint and default model together.

After setting a default model, close Settings, choose New Chat in the sidebar, type a message, and send it.

Use a large GGUF without splitting it

You do not need to split a large GGUF before using it with Naidan. Even a single file tens of gigabytes in size can be imported from your device or selected from its Hugging Face repository. For a repository file, Naidan downloads it to the chosen storage, then reads the saved GGUF in bounded chunks.

If your model does not run

Check Inference settings first, then confirm that the GGUF and any companion mmproj match the model you intend to use. A completed download only means the files are available; it does not prove that this browser can run the model. Standalone includes WebGPU with JSPI for wasm32 or wasm64, without a CPU or Asyncify fallback.