Nous Research Adds One-Click Local Model Setup to Hermes Desktop

Hermes Desktop is the MIT-licensed, free build of the open-source Hermes Agent. It runs on macOS 12 or later, Windows 10/11, and any Linux distribution, and requires no account for local models. The one-click flow appears automatically on first launch and can be reached later under Settings → Providers → Local Models, the report said. Hermes manages the inference engine itself: it fetches an official llama.cpp build matched to the hardware, verifies it, and keeps it updated. The supported backends are CUDA, Metal, Vulkan, HIP, and CPU. The pinned release tag is stored in the local_runtime block of config.yaml, which the desktop UI writes for the user and which headless users can set by hand.

Before any download, every model in the catalog is evaluated against the specific machine. A color-coded memory-fit verdict marks each row: green means the model runs entirely in GPU memory; amber means it spills into system RAM and runs slower; red means it is too large for this machine. The rows also display the starting and maximum context windows and the download size of the build selected for the hardware. Hermes chooses the highest-quality quantization that runs fully on the GPU, with a hard floor at 4-bit; models that cannot fit at 4-bit remain visible with the reason attached, so users can see what additional VRAM would buy.

The application exposes no manual controls for context length, GPU layers, or quantization. Models start with a context window that fully fits the GPU and grow toward their native maximum as conversation needs room, with every recommended model guaranteed at least a 64K window, according to the report. When a model exceeds GPU memory, Hermes places the overflow into system RAM in the order that hurts least: expert weights first, and never the attention cache. Conversation compression only engages once the model hits its maximum window. Idle models are unloaded after 15 minutes and reloaded on the next message.