HermitUI is a lightweight, modern, and highly responsive single-page web interface for interacting with local AI models. Built entirely in a single file, it requires no build steps, backend, or installation.
By design, it prioritizes your privacy—no chat history is saved across sessions.
Accepts direct .gguf links, Hugging Face /blob/ page URLs, or hf:user/repo/file.gguf. The model is downloaded straight into memory and re-downloaded next session — nothing is ever stored.
Status: Unloaded
Auto uses the model's embedded template when present, else infers from its architecture.
Larger = more history but more memory/slower load (the KV cache grows with it — if the model fails to load, lower this). 0 = use the model's built-in default. Applied when the model is loaded.
Upper bound on tokens generated per reply. Applied immediately (no reload needed). Must fit within the context window.
Note: On some systems (especially integrated graphics) WebGPU may be slower than CPU. Uncheck to force CPU-only mode.
🐛 Debug output (engine load, model metadata, chosen template, the prompt sent, tokens/sec) streams to the drop-down console at the top of the screen — click the 🐛 button in the header.
⚠️
Privacy Warning: You are connecting to a cloud provider. Your chat data will leave your local machine.
👁️ vision
Advanced parameters
⚠️Privacy Warning: You are using a cloud provider (). Your chat data leaves your local machine.
🔗This link wants to load from for in-browser inference.