Guide

Hermes Agent + Kimi K3 through OpenRouter

Q: Does Hermes Agent support Kimi K3?

Yes. Hermes Agent lists moonshotai/kimi-k3 through its OpenRouter provider. Configure OPENROUTER_API_KEY and select that exact model ID.

Q: Should I switch an existing Kimi session to K3?

Start a new Kimi K3 session rather than switching an existing thread. Moonshot says the model expects its thinking history to remain intact.

Hermes already lists Kimi K3 in its OpenRouter catalog. Use that verified path for launch day; the direct Kimi provider still lists K2-family models.

What works today

Hermes' current catalog includes moonshotai/kimi-k3 under OpenRouter and Nous Portal. The most reproducible BYOK path is OpenRouter: add an OpenRouter key, choose OpenRouter in hermes model, and select that exact provider-prefixed model ID. The current direct kimi-coding and moonshotcatalogs still stop at K2-family models, so do not tell Hermes that a different Kimi entry is K3.

Configure OpenRouter

Store the key in Hermes' private environment file rather than a shell-history command or committed YAML. Open ~/.hermes/.env in a text editor and add this file content:

OPENROUTER_API_KEY=your-key-here

Restrict the file, then use the interactive live model picker:

chmod 600 ~/.hermes/.env
hermes model

Create a key at OpenRouter.
Add it as OPENROUTER_API_KEY in ~/.hermes/.env.
Run hermes model, choose OpenRouter, and select moonshotai/kimi-k3.
If it is absent, update Hermes and refresh the live catalog. Do not rename a K2 model to K3.
Start a new session and test one read-only tool before enabling broader actions.

The model ID and provider lists are visible in Hermes' official model catalog. OpenRouter is a third-party gateway with its own billing and data handling, not Moonshot's first-party API.

Why K3 fits Hermes

Long agent runs: a 1M-token window leaves room for plans, tool results, and large code context.
Terminal and repository work: Moonshot specifically highlights long-horizon coding and navigation.
Visual tasks: native vision can interpret screenshots used in browser and UI workflows.
Cost-aware research: long context still needs budgets; check OpenRouter's live model pricing before a large run.

Important operating rules

Moonshot's Kimi K3 technical post says thinking history must be preserved. Do not swap K3 into the middle of a K2.6, Claude, or GPT thread. It also notes excessive proactiveness as a limitation. For Hermes, keep shell, browser, and external-write tools behind explicit approvals until you have evaluated the model on your own tasks.

API access is not the same as local weights

The hosted API launched July 17, 2026. Moonshot says the full weights arrive by July 27 and recommends 64 or more accelerators for deployment. A workstation running local Ollama is therefore not a realistic K3 target for most users. If you need local inference, use the Hermes + Ollama guide with a smaller model.

Validate before a long run

Confirm the selected provider and exact model name in hermes model.
Test a short answer, a structured tool call, and a follow-up that depends on earlier context.
Watch usage in OpenRouter's dashboard; 1M context can still produce a large bill.
Keep an existing stable model configured as a fallback while K3 availability settles.

Frequently asked questions

Does Hermes Agent support Kimi K3?

Yes, through OpenRouter. Add OPENROUTER_API_KEY, then select moonshotai/kimi-k3 from the OpenRouter catalog in hermes model.

Should I switch an existing Kimi session to K3?

Start a new session. Moonshot warns that K3 expects its thinking history to be preserved, so changing models midway can degrade behavior.