Keep exploring · more guides
Hermes Agent just got a free brain upgrade, and it runs on your own computer.
Imagine opening your laptop and finding a giant AI model already sitting there waiting for you.
No account. No key. No internet needed.
It reads pictures, it uses tools, and it thinks before it answers.
Your own private AI employee, working for free, forever.
There is also one setting almost everyone gets wrong, and I will show you how to fix it in ten seconds.
You do not own any of it.
Every question goes to somebody else's computer, and every answer comes back on their terms.
What you're looking at: the rented brain. Today we move in.
It is four commands, and every one of them is on this page.
If you can type a sentence into a text box, you can do the whole build.
A model that only chats is a toy. A model that uses tools is an employee.
What you're looking at: the four things that make this one worth the disk space.
MLX is Apple's own way of running AI on the chip inside your Mac.
Same model, same answers, built for the machine you already own.
What you're looking at: don't drop the mlx. That is the whole tip.
Three green ticks and you are good to go.
What you're looking at: the three-second sanity check that saves you an hour.
Open Terminal and run this. If your Ollama is old, the download fails with a code nobody explains.
ollama --version
What you're watching: my real session. Version 0.32.5 looked fine, then the pull died on a 412 — "requires a newer version of Ollama". The MLX builds are too new for older Ollama to even read.
Download Ollama again from ollama.com, drag it into Applications, replace the old one, reopen it.
What you're watching: 0.32.5 became 0.32.14 on my machine — and the new build ships one-command wiring for Claude Code, OpenCode, Hermes and a dozen more. We use the Hermes one at §11.
It did not touch mine. Every model I already had was still listed straight after the upgrade.
Your models live in a separate folder from the app, so replacing the app leaves them alone.
Run df -h ~ and read the "Avail" column.
I had 15 GB free and needed 18. Three deletions later I had 63.
What you're looking at: four models I had already tested and moved on from months ago. Deleting them took three seconds.
This is the download. Eighteen gigabytes, once, and then it is yours forever.
ollama pull qwen3.8:27b-mlx
What you're watching: the real pull finishing, then ollama list showing it sitting on my disk at 18 GB. From here it works with the wifi switched off.
Your Mac has to lift 18 GB off the disk before it can think at all. Mine took about 35 seconds.
Every answer after that comes back straight away.
What you're watching: the real first reply, then a warm one — and the number I measured on an M4 Max: 24.9 tokens a second.
That is the load, not the thinking, and it only happens once per session.
Ask it a second question and the answer starts immediately, because the model is already awake.
A number on its own means nothing, so here it is against a human.
What you're looking at: the answer unrolls faster than you can read it, so it feels like it was already there.
Ollama holds the model. Hermes is the agent. This joins them.
ollama launch hermes --model qwen3.8
What you're looking at: the join. Before this, connecting a local model to an agent meant editing files and guessing.
Hermes keeps a folder per personality — a profile. Make one, drop this file in, and you can point it anywhere.
# Hermes "qwen38" profile — Qwen 3.8 27B (MLX 4-bit) via Ollama. # 100% offline. Apple-Metal MLX build. ~25 tok/s warm on an M4 Max. # Run it with: hermes -p qwen38 model: default: qwen3.8:27b-mlx provider: ollama-launch base_url: http://127.0.0.1:11434/v1 context_length: 8192 # ← the setting from §23. keep it small. ollama_num_ctx: 8192 # ← this is the one that eats your RAM. providers: ollama-launch: api: http://127.0.0.1:11434/v1 default_model: qwen3.8:27b-mlx models: - qwen3.8:27b-mlx name: Ollama (local, offline) # nothing in the cloud to fall back to — that is the point fallback_providers: [] toolsets: - hermes-cli agent: max_turns: 60 gateway_timeout: 1800 # local models think slower than the cloud api_max_retries: 2 tool_use_enforcement: 'off' environment_probe: false terminal: backend: local cwd: ~/.hermes/profiles/qwen38/workspace # then: hermes profile list → your profile appears in the table # hermes -p qwen38 → you are talking to your own machine
What you're watching: the real profile file, scrolling. Watch ollama_num_ctx — that is the setting we come back to at §23.
Same model in both cases. Completely different day.
What you're looking at: the four things a bare chat box cannot give you.
If you handle client files, medical notes or legal work, this is the version of AI you are actually allowed to use.
Read every file in this folder, give me a one-line summary of each, then tell me which three need my attention first.
What you're looking at: nothing crosses the dotted line. That is the whole privacy story, in one picture.
A dashboard, a chart, a messy spreadsheet, a photo of the whiteboard after a meeting.
What am I looking at, and what should I do about it?
What you're looking at: most free local models cannot do this at all. This one reads images out of the box.
It decides, it runs something, it reads what came back, and it goes round again until the job is done.
What you're looking at: the loop that turns a chatbot into a worker.
This whole page is the answer to that. The everyday work runs on a free model on your own machine, with no meter at all.
Free APIs slot in alongside it, and for the hard jobs Agent OS drives the CLIs you already pay for — your Claude subscription already includes the Claude CLI, so you are not paying twice. Inside the Boardroom there are full token-efficiency tutorials on top.
Thinking is on by default. For simple questions that is just slow.
What you're looking at: the dial nobody touches. Move it and the model feels like a different product.
Drop in a whole contract and ask what is hiding in clause forty.
What you're looking at: it can spot things across the whole document instead of one page at a time. (Careful — see §23 before you actually turn this all the way up.)
What you're looking at: same work, two completely different relationships with it.
Wrong: "Local models are rubbish compared to the big ones."
Right: That was true two years ago. This one is 27 billion parameters, reads images and uses tools — the gap closed while people were repeating the old line.
Wrong: "This is too technical for me."
Right: It is four commands — check, upgrade, download, launch. Every one of them is printed on this page.
Wrong: "I already pay for an AI, so there is no point."
Right: Keep it. This is a second worker who never sends an invoice — free one for the 90% that is grunt work, paid one for the hard 10%.
Members post their wins every day — agency owners, ecom founders, course creators and solo operators across 38 countries. Real businesses, real numbers, in their own words.
Read the 158-page wins doc →You can wire all of this yourself from this page. Or get the whole thing done, inside the Agent Operating System.
Eighteen gigabytes sits in your memory while the model is awake.
I set it up greedy, left everything else running, and my whole Mac crawled.
What you're looking at: same model, same Mac. The only difference is two settings — and they are in the next section.
The model reserves memory for your whole context number the moment it wakes up, whether you use it or not.
So drop it, and tell the model to let go when you stop working.
What you're looking at: put the context back up only when you actually need to feed it a book.
No — the model really does hold it. Your Mac just has to be able to afford it.
Keep it small for daily work and raise it for the one job that needs it. That is a choice you only get to make because the model is yours.
Two commands. One puts it to sleep, one proves it worked.
ollama stop qwen3.8:27b-mlx && ollama ps
What you're watching: the real unload. The table comes back empty and the RAM is yours again. The model is still on your disk — the next question wakes it up.
A smaller model that runs beautifully beats a big one that makes you hate your computer.
What you're looking at: my own verdict after running it on a 36 GB machine. I'm in row three, and row three needs the settings.
Disk space and a bit of electricity. That is the whole bill.
What you're looking at: no key, no account, no meter. Every question you ask for the rest of the year adds nothing.
Everything on this page comes from the official pages. Click any of them and check me.
One 18 GB download and it is yours. No key, no account, nobody to switch it off.
About three times faster than you read out loud, measured warm on an M4 Max.
Turn the wifi off and it still answers. Client files stay client files.
ollama launch hermes --model qwen3.8
The two settings that decide whether this is a joy or a nightmare.
Turn it down and the same model feels twice as fast.
This page gives you one free brain on your own machine. The Agent Operating System gives you everything I built around it — the local model already wired in, Hermes profiles already made, and every AI you pay for and every AI that is free sitting in one dashboard, sharing one memory.
Four coaching calls a week where we set it up on your machine with you. Daily tutorials as new models land. 4,000+ founders across 38 countries running it right now.
Readers bookmark this page and never open the Terminal. Operators install it tonight and wake up with an AI employee that works for free.
Get the Agent OS → Inside the AI Profit Boardroom · skool.com/ai-profit-lab
