Run an AI model in your browser
To try a local LLM in your browser, you can run a chat model on your device without starting a separate inference server. The browser loads the model and performs the generation. Start with the llama.cpp choice if you have a GGUF model or want to choose one from the built-in catalog.
Model files may need to be downloaded first. That download contacts the model distributor; afterward, generation uses the selected browser runtime. Browser and hardware support determine which models will run and how fast they respond.
Choose the browser runtime that fits your setup
llama.cpp in the browser
Choose a suggested GGUF model, inspect a Hugging Face repository, or import your own GGUF file or model folder. The model manager shows downloaded files and lets you choose browser storage or a supported device folder as the destination.
View setup guideTransformers.js
Open the Transformers.js manager and download a supported ONNX model. When the download completes, Naidan starts loading it automatically; the downloaded list offers Load for a complete model or Resume for an incomplete one. Its model files are stored in the browser’s OPFS. This option is unavailable in Standalone because local-file browser restrictions affect its workers and WebAssembly.
View setup guideBrowser-provided model
Choose Browser-provided and read the availability status. If the browser offers a model that needs preparation, use Prepare browser-provided model. The browser manages that model, so its availability and download behavior depend on the browser.
View setup guideStart a chat with a GGUF model
Select llama.cpp
At the first-run setup, choose llama.cpp under In-Browser AI. If you have already started using Naidan, open the llama.cpp browser tab in Settings to manage models, then choose llama.cpp browser (Experimental) as the API Provider on the Connection tab.
Prepare the model files
Use the Model catalog to inspect a suggested model and its size before Download, or import a GGUF file or model folder you already have. For a Hugging Face repository, check the file plan and quantization before downloading. Local-only cannot download from Hugging Face; import a GGUF obtained separately, or download in a Hosted edition that permits the required model file hosts. Large models take storage and memory.
Select it for conversation
Pick the imported model in Default Model during setup, or choose it in the Connection settings and Save Changes. Existing and new conversations that inherit the global model use this default; a chat or group can choose another model.
Try a short message
After first-run setup, select Get Started. If you changed an existing setup, close Settings after Save Changes and select New Chat. Send a short text prompt. If the model is missing, Naidan offers a path back to model preparation while keeping your draft.
Before using images or other media
To attach an image to a llama.cpp chat, you need a matching mmproj GGUF, the model’s companion file for image input. The current llama.cpp chat input does not support audio or video. Generating an image or speech uses separate model features and setup.
Standalone includes browser llama.cpp through its embedded WebGPU runtime, without a CPU fallback. The hosted edition has more execution profiles. Check the in-app availability notice on your browser and device. Transformers.js is a hosted-only choice.
Downloaded models and chat history have different storage controls. Removing browser site data can remove cached models; choosing a device model folder depends on browser permission and folder access.