Kimi K3.1 Pro Hermes setup takes about five minutes: put your Moonshot key in ~/.hermes/.env as KIMI_API_KEY, then run hermes chat --provider kimi-coding --model kimi-k3.1-pro. Hermes has Kimi built in as a first-class provider, so there is no custom endpoint to wire up.
📺 Watch: Kimi K3.1 Pro is INSANE!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
I run Hermes daily inside my Agent OS with a profile per model. Kimi K3.1 Pro just earned its own profile, because Moonshot has it less than a point behind Claude Opus 5.5 on Terminal-Bench 4.0 at about half the price.
This guide is the practical part. The key, the test run, two Kimi profiles, the one setting you must change, and which agent jobs I route to Kimi and which I keep elsewhere.
If you want my full Agent OS and the profiles I use every day, it is inside the AI Profit Boardroom for $69/mo, along with a 30-day AI agent roadmap and four coaching calls a week.
Before you start
You need two things.
- Hermes Agent installed. Any recent version lists Kimi / Moonshot as a provider.
- A Kimi API account. Sign up on Moonshot's platform at platform.moonshot.ai and create an API key.
One thing trips people up. Moonshot only unlocks K3.1 Pro once you have topped up at least $1. Your rate limits also grow with how much you have topped up in total, so a brand new account will hit limits sooner on long agent runs.
If you are new to the model itself, my Kimi K3.1 Pro breakdown covers the benchmarks, pricing and open weights. This page is all setup.
Step 1: add your Moonshot key to Hermes
Hermes reads the Kimi key from your env file. Open ~/.hermes/.env and add one line:
KIMI_API_KEY=your-key-here
That is the key for the kimi-coding provider, which Hermes points at Moonshot's international endpoint, https://api.moonshot.ai/v1, by default. Hermes also accepts KIMI_CODING_API_KEY as an alias, and you can override the endpoint with KIMI_BASE_URL if you ever need to.
If your account is on Moonshot's China platform instead, Hermes has a separate kimi-coding-cn provider that reads KIMI_CN_API_KEY.
📺 Watch: Kimi K3 + Hermes AI Agent is INSANE!
Step 2: test Kimi K3.1 Pro in Hermes
Before building profiles, do a one-off test so you know the key works:
hermes chat --provider kimi-coding --model kimi-k3.1-pro
Give it a small real task, like reading a folder and summarising what is in it. You should see it think, call tools and answer.
Want the faster version? Swap the model ID:
hermes chat --provider kimi-coding --model kimi-k3.1-pro-highspeed
Moonshot says highspeed runs 2 to 3 times faster. It costs double: $4 per million input tokens and $20 per million output, against $2 and $10 for the standard model.
📺 Watch: Kimi K2.7 + Hermes Subagents
Step 3: set the reasoning effort
This is the setting most people miss.
Kimi K3.1 Pro always thinks. You cannot switch thinking off. Moonshot's API takes three effort levels, low, high and max, and max is the default.
Hermes, on the other hand, sends medium when you have not set an effort. Its docs point out that on kimi-k3, max uses about three times the reasoning tokens and latency of medium. So pick a level on purpose instead of leaving it to defaults. Inside a Hermes chat:
/reasoning high --global
The --global flag saves it to that profile's config.yaml, so it sticks. Use high for hard coding and research jobs, and low for quick tasks where speed matters more.
Not sure which level fits your workflows? A free AI strategy session is a fast way to get that mapped out.
Step 4: give Kimi K3.1 Pro Hermes profiles of its own
I never swap models inside one profile. Each model gets its own profile, with its own config, memory and state. That way Kimi's memory does not get mixed up with Claude's, and I can run them side by side.
Create the main Kimi profile:
hermes profile create kimipro
Hermes creates the profile plus a kimipro command alias. Then run setup and pick Kimi / Moonshot as the provider and kimi-k3.1-pro as the model:
kimipro setup
If you prefer to set it directly:
kimipro config set model.provider kimi-coding
kimipro config set model.default kimi-k3.1-pro
Now make a second, faster profile for quick jobs:
hermes profile create kimifast
kimifast config set model.provider kimi-coding
kimifast config set model.default kimi-k3.1-pro-highspeed
Open each one with kimipro chat or kimifast chat, and set the effort once in each: high in kimipro, low in kimifast.
I name them kimipro and kimifast rather than kimi so they never clash with any other kimi command on your machine.
Which agent jobs to route to Kimi K3.1 Pro
Here is how I split the work, based on where Moonshot's own benchmarks say K3.1 Pro is strong and weak.
| Agent job | Profile | Why |
|---|---|---|
| Long coding and terminal tasks | kimipro, high effort | 65.7% on Terminal-Bench 4.0, against 66.4% for Opus 5.5, at about half the price |
| Research write-ups and fact-heavy reports | kimipro, high effort | Hard hallucination rate down to 8.3 with tools, from 23.4 on K3 |
| Spreadsheets, finance and office documents | kimipro | Moonshot claims a state-of-the-art 64.4 on OfficeQA Pro v2 |
| 3D scenes and game assets | kimipro | 72.8 on BenchCAD without tools, next to Opus 5.5's 73.0 |
| Quick summaries, triage, inbox sorting | kimifast, low effort | Speed matters more than depth |
| Live web research | Another profile for now | Moonshot says its built-in web search is being updated and not recommended for production yet |
| Security work | Another profile | Moonshot deliberately cut cyber training data and says it trails frontier closed models here |
The pattern is simple. Kimi handles the heavy, long-running jobs where cost adds up fast. Claude keeps the jobs where it still leads, like 3D work with tools.
Long context is the other reason to route to Kimi. K3.1 Pro has a 1M-token window, so an agent can load a big codebase or a stack of documents in one go. Moonshot also caches repeated prompt prefixes automatically, at $0.20 per million cached input tokens, which helps on agent loops that resend the same system prompt every turn.
Before you move a job over for good, test it. Run the same task through your Kimi profile and your current profile and compare. That is exactly how I built Goldie Bench.
Getting more out of it
A few habits help on long agent runs.
- Keep the system prompt stable. The cache works on repeated prefixes, so do not keep editing the top of your prompt mid-run.
- Use structured output for pipelines. Moonshot supports strict JSON schemas, which keeps hand-offs between agents clean.
- Turn on Zero Data Retention on your Kimi API account if you are sending client work through it. Moonshot offers it to API customers.
- Watch your tier. If a long run stalls on rate limits, it is usually the top-up tier, not Hermes.
Once the open weights land, which Moonshot says will be by November 4, 2026, more hosts will be able to serve it. For now, Moonshot's own API is the route Hermes supports directly.
If you ran the previous model, my Kimi K3 Hermes Agent guide covers the K3 setup, and Kimi K3 vs Fable 5 puts K3 head to head with Claude. For how all these profiles fit into one system, read my Agent OS guide.
Want someone to look at your agent setup and tell you which jobs to move first? Book a free AI strategy session.
FAQ
Does Hermes Agent support Kimi K3.1 Pro?
Yes. Hermes has a built-in kimi-coding provider for Moonshot's API. Add KIMI_API_KEY to ~/.hermes/.env and use the model ID kimi-k3.1-pro.
Why can I not use K3.1 Pro on a new Kimi account?
Check your Kimi account balance first. Moonshot only unlocks K3.1 Pro after a top-up of at least $1.
Should I use Kimi K3.1 Pro or Highspeed in Hermes?
Use the standard model for deep coding and research, and highspeed for quick jobs. Highspeed is 2 to 3 times faster but costs double per token.
Can I run Kimi K3.1 Pro locally in Hermes?
Not yet. Moonshot says the weights come out by November 4, 2026, and even then it is a 3T-class model that needs serious hardware. The API is the practical route.
Ready to run Kimi K3.1 Pro alongside Claude in your own Agent OS? Join the AI Profit Boardroom, or grab a free AI strategy session to plan your setup.
