> ## Content Index
> Fetch the complete content index at: https://globalfeed.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Nous Research's Hermes Desktop sets up a local model in one click: reads the hardware, picks the build, ships llama.cpp in the package
- URL: https://globalfeed.ai/en/hermes-desktop-now-sets-up-local-models-in-one-click/
- Published: 2026-09-06T13:43:48.000Z
- Updated: 2026-09-06T13:43:48.000Z
- Description: Hermes Desktop, the open-source agent app from Nous Research, now handles local model setup in a single click. The app reads the machine's hardware, rates every model in its catalog against that machine, downloads the one that fits and configures the llama.cpp runtime itself. The feature appears on
- Author: GlobalFeed Editor
- Tags: Nous Research, Hermes Desktop, local ai, agents, open source, x-NousResearch, dil-en, video, x-gitti

**Hermes Desktop**, the desktop app for Nous Research's open-source Hermes Agent framework, now sets up a local model in one click. Per the company's [3 September post on X](https://x.com/NousResearch/status/2095602995874410664?ref=globalfeed.ai), the app reads the hardware, picks the best model, downloads it and configures the runtime. Setup is offered on first launch and sits under Providers in Settings, alongside Nous Portal, OpenRouter, OpenAI and custom endpoints.

## How it picks a model

Per the [local models documentation](https://hermes-agent.nousresearch.com/docs/user-guide/local-models?ref=globalfeed.ai), every catalog model is rated against the machine before download: **green** runs entirely in GPU memory, **yellow** spills into system RAM and slows down, **red** does not fit. Each entry lists download size, starting and maximum context window.

The documented rule: Hermes picks the highest-quality build that runs fully on the GPU; machines with less memory get a smaller build of the same model. The floor is 4-bit quantization, below which Nous Research counts the quality loss too high.

## Runtime and context

The package ships a hardware-matched **llama.cpp** build of a few hundred megabytes, with CUDA, Metal, Vulkan, HIP and CPU backends, pinned in the `local_runtime` block of `config.yaml`; the app verifies updates itself.

Recommended models guarantee at least **64K** of context, starting at a size that fits the GPU and growing toward the model's native maximum as the conversation lengthens. On overflow, expert weights move to RAM first; the attention cache never spills. Compression starts only at that maximum, and an idle model is unloaded after 15 minutes.

## Platforms and license

The app runs on macOS 12 and later, Windows 10 and 11, and Linux distributions. It is MIT-licensed, free and open source; local models need no account.

## For the corporate reader

Setting up an in-house agent that keeps data inside gets easier: model, quantization and context size are now the app's job. Two limits: the "best model" choice rests on Nous Research's catalog and criteria, so organizations should validate for themselves, and the documentation does not list the catalog's models.

## Sources

- **Announcement:** Nous Research on X, 3 September: [x.com/NousResearch/status/2095602995874410664](https://x.com/NousResearch/status/2095602995874410664?ref=globalfeed.ai)
- **Documentation:** Local Models: [hermes-agent.nousresearch.com/docs/user-guide/local-models](https://hermes-agent.nousresearch.com/docs/user-guide/local-models?ref=globalfeed.ai)
- **Download:** [hermes-agent.nousresearch.com/desktop](https://hermes-agent.nousresearch.com/desktop?ref=globalfeed.ai)
- **Source code:** [github.com/NousResearch/hermes-agent](https://github.com/NousResearch/hermes-agent?ref=globalfeed.ai)