This setup is two files. An OpenRouter API key in .env and a config.yaml. That gets you the model ranked #1 for frontend code on the Frontend Code Arena at the time of recording, inside the most popular agent on OpenRouter.
Why this pairing keeps coming up
Kimi K3 is Moonshot’s open-weight model with a 1-million-token context window. Hermes Agent is the most used agent on OpenRouter, and because OpenRouter routes to basically any model behind one API, dropping Kimi K3 into it is a config change rather than a project.
If you want to run Kimi K3 without transcribing settings from a screen recording, I put the exact copy-paste configs here:
github.com/ravsau/ai-tutorials/hermes-kimi-k3-openrouter
The video walks through the whole Kimi K3 setup on camera, so you can check your version against a working one.
Before you get too excited
I made a separate video on this model called “Kimi K3 Beat Claude and GPT — But There’s a Catch”, and the catch belongs in this post too. People testing it on real work found it slow and extremely token-hungry.
That second part is the one to sit with. Price per token is not cost per completed task. A model can look cheap on the pricing page and still burn through so many tokens per job that the total bill surprises you. Benchmarks measure quality of output. Your invoice measures something else.
The sensible way to test it
Set it up with the repo configs, point it at one genuine task from your own backlog, and pay attention to elapsed time and tokens used. Twenty minutes of honest testing beats any leaderboard screenshot.
Watch the full Hermes Agent OpenRouter walkthrough in the video above, grab the configs from the GitHub repo, and subscribe if you want the next setup tested the same way, catch included.