Hermes + Qwen 3.8 Max — a 2.4 trillion parameter brain in one config file.
You can now point your AI agent at a two point four trillion parameter brain, and it takes about two minutes.
That means the biggest model Alibaba has ever built is doing your work instead of sitting behind a chat box on someone else's website.
It holds a million tokens at once, so it keeps the whole job in its head from start to finish.
I wired it up this morning and gave it two builds — both worked first time, and one of them added something to the artwork I never asked for.
Today: the exact setup, the builds it produced, and the one mistake that stops most people getting it running at all.
Stick with me to the end. Let's get into it.
Here's the pairing in one breath.
Hermes is the agent — the thing that plans, runs commands, writes files and finishes work on your machine.
Qwen 3.8 Max is Alibaba's largest model: 2.4 trillion parameters, about 95 billion active at a time, with a one-million-token memory.
On its own it's a very large brain with no hands.
Inside Hermes it becomes something that does your work while you're somewhere else.
2.4Tparameters behind your agent
1Mtoken context window
2 minto wire it in
2/2builds worked first try
II ────── Why bother
A giant brain in a chat box is a very expensive toy.
Most people meet a new model the same way. They open the company's website, type something, read the answer, say "that's clever", and close the tab.
Nothing happened. No file was written. No job got finished.
An agent changes what the model is. Same brain, but now it can read your files, run commands, check its own work and keep going without you sitting there.
That's why the pairing matters more than the model:
The size buys you patience. A million tokens means it can hold an entire job — every file, every result, every earlier decision — instead of forgetting what it did twenty steps ago.
The agent buys you hands. Hermes turns "here's how you'd do that" into the thing actually being done.
The profile buys you freedom. Once the model is one config file, swapping brains is a flag, not a rebuild. Today it's Qwen. Next month it's whatever wins.
The model brings the memory. The agent brings the hands. Neither is much use alone.
"Do I need the biggest model for everyday work?"
Honestly, no — most of what you do runs fine on something small and cheap.
Reach for a brain this size when the job is long, tangled, or spread across a lot of files, which is exactly where smaller models start losing the thread.
The point of a profile is that you can keep both and pick per job.
✦
III ────── The setup
One folder. One file. Two minutes.
A Hermes profile is a named folder holding a config file. That's all it is.
Step 1 — make the folder. Create ~/.hermes/profiles/qwen-3-8/.
Step 2 — write the config. Inside it, a config.yaml:
Step 3 — give it a key. Alongside it, an auth.json holding your provider key. If you already run other cloud profiles, copy the credential block across.
Step 4 — check it took, before you trust it.
hermes -p qwen-3-8 status
You want to see Model: qwen/qwen3.8-max and the auth file pointing at your new profile. If it shows a different model, you're not running what you think you are.
Step 5 — use it.
hermes -p qwen-3-8 -z "build me a thing" --yolo --accept-hooks
One flag switches your agent's brain. Everything else about your setup stays exactly as it was.
"What if I don't use Hermes?"
The shape is the same anywhere. Any agent that lets you set a custom model name and endpoint can point at this — the config file just has a different name.
What you're really doing is separating "which brain" from "which tool", and that's worth doing whatever you run.
✦
IV ────── The mistake that stops people
The model name is right. The route is dead.
This is where I lost time, so you don't have to.
My first attempt pointed at a local gateway that advertised the model. The name in its list was correct. Everything looked right.
Every request failed — because that gateway was still pinned to the model's preview version, an ID the company retired the moment the finished version shipped. It didn't matter what model name I sent. The route itself was stale.
Two lessons worth keeping:
1 — A model appearing in a list is not proof it runs. Send one real message before you believe anything. Mine now answers "QWEN 3.8 MAX LIVE" on demand, which takes ten seconds to check.
2 — Made-up provider names fail silently until they don't. I invented a provider label in the config and Hermes rejected it outright: unknown provider. Use a provider name your agent already knows, and point it wherever you like with the base URL. The provider field names a protocol, not your nickname for the endpoint.
Both mistakes look identical from the outside: a correct-looking config that does nothing. Ten seconds of testing tells them apart.
Two configs that look the same in a text editor. Only one of them answers.
✦
V ────── What it built
Two jobs. One prompt each. Both landed first time.
Setup means nothing without proof, so I gave the pairing two builds and let the agent do everything — plan it, write the file, save it to disk.
No follow-ups. No corrections. Nothing fixed by hand afterwards. I checked both in a real browser rather than taking the model's word for it.
The Jade Engine is the harder test — real 3D, interlocking parts, and a number on screen that has to actually move. It hit all of it.
But the Ink Garden is the one that made me sit up. I asked for ink-wash branches, blossoms and a seal stamp. It also added Chinese characters in the corner and matched them inside the seal — something I never mentioned. That's a model with real cultural range rather than a translation layer bolted on top.
Being straight about what this proves: two builds show it's competent and follows a brief closely. It doesn't tell you where it ranks against the frontier leaders. That takes proper testing, and anyone claiming a ranking from a couple of demos is guessing.
Then we handed the same brain a camera. Higgsfield's studio speaks MCP, and its sign-in turns out to be portable — copy one mcp-tokens folder into the qwen-3-8 profile, add four config lines, and the giant engine suddenly owns 90 image, video and audio tools. No second login. Same chat.
One brief per set, nobody watching. It explored the model catalogue itself, chose the engines, rendered, and downloaded every file. For the book set it even ran the reference-image upload flow so the shots show my real cover — not an AI's guess at it.
Web apps, cinema, print-grade stills — one profile, one flag, zero babysitting. The full studio wiring (and the two-file token move) is in the Higgsfield MCP guide.
No step in the middle needs you. That's the whole difference between a model and an agent.
"Is it fast?"
Not especially — each build took several minutes of thinking before any code appeared.
That's the trade with big models: you're buying the ability to finish something long and complicated, not a quick answer.
Give it work you'd walk away from, and the wait stops mattering.
✦
VI ────── My story · why this matters
I was you. Then the brain became a swappable part.
Before
Every big launch cost me a day.
New model, new tool, new setup, new habits to learn.
By the time I'd wired it in, something else had launched and I was behind again.
And everything I built sat on top of one company's decisions about price and availability.
Then I made the agent permanent and the model a config file.
After
The largest model anyone shipped this week was running in my setup before lunch.
One folder, one file, one flag.
Every agent, note and workflow I already had carried straight over.
If it wins on real work, it stays. If not, I've lost an afternoon and learned something.
You can have this too. Build the system once, then let the models come to you.
VII ────── The receipts
Real people. Real builds. Inside the Boardroom right now.
The members swapping engines the week they ship — agency owners, ecom founders, course creators.
4,000+ founders inside AIPB
258 real wins documented
400K YouTube subscribers
38 countries · live members
163K X followers
Members post their wins as they happen — first agents shipped, workflows automated, whole businesses run from one dashboard. All collected in one doc you can read right now.
Two minutes. One folder, one file, one test command.
So here's the deal.
Make the profile today and give it one real job from your own week — something you'd otherwise do yourself.
Not a riddle. Actual work.
The people reading launch threads this week will still be reading them next week.
The people who wired it in have finished work to look at.
Be one of those people. One profile. Today.
✦
VIII ────── Old way vs new way
Most people visit models. Operators plug them in.
The old way: a huge model launches, you open its website, type a few things, feel impressed, and close the tab.
Nothing in your business changed.
Next month a bigger one launches and you repeat the ritual.
The new way: the agent is permanent and the brain is a part you swap.
A launch is a folder and a config file, and everything you've already built inherits it instantly.
You test the new brain against the old one on real work, keep the winner, and move on with your week.
✦ ✦ ✦
Get the whole operating system
Want this pre-wired, not DIY?
Every engine in one dashboard, swappable, with the profiles already built — that's the Agent OS. Join the AI Profit Boardroom and you get:
The Giant Engine pre-wired — Hermes with the Qwen profile ready to run
The full Agent OS — every major engine in one dashboard, new ones added the week they ship
Free local models — the everyday 90% of work at $0 on your own machine
Agent Kanban — Planner → Builder → Reviewer teams that ship while you sleep
The memory vault — an Obsidian brain your agents actually read
Model-routing playbooks — which engine for which job, kept current as the field moves
5 live coaching calls a week — get unblocked by people on your exact setup
You're not buying a tool. You're getting the operating system I run a seven-figure business on — and it absorbs launches like this one without breaking stride.
Inside the AI Profit Boardroom · skool.com/ai-profit-lab
Set up in an afternoon · used in 38 countries · new tools added every week
X ────── Three beliefs to drop
What's actually stopping you isn't technical.
"Setting up a custom model sounds like developer work."
It's a folder with a short text file in it and one command to check it worked.
If you can make a folder and paste eight lines, you've done the hard part.
Everything after that is just talking to your agent like you already do.
"A model that big must be overkill for me."
For a quick question, completely. For the jobs you keep putting off because they're too tangled to start, that size is exactly the point.
Big context is what lets an agent work through something long without losing the thread — and those are the jobs worth automating.
"There'll be a better model next week, so why bother now?"
There will be — and that's the argument for doing this, not against it.
Once your agent treats the brain as a swappable part, next week's model is another two-minute profile instead of another rebuild.
You're not betting on one model. You're building the socket they all plug into.
✦
XI ────── The SOP
Do it properly — five steps.
1 · Make the profile. A folder under your agent's profiles directory with the config from section III. Two minutes.
2 · Verify before you trust it. Run the status command and confirm the model and auth file are the ones you expect. This is the single most valuable habit in this guide — a wrong label costs you hours of confusion.
3 · Send one real message. Ask it to reply with a specific phrase. If that comes back, the route works. If it doesn't, fix the route before blaming the model.
4 · Give it a real job, not a quiz. A whole build, a multi-step task, something with files in it. Puzzle answers tell you nothing about whether it can finish your work.
5 · Use the context you're paying for. A million tokens means whole codebases and whole research piles fit in one go. Most people never change their habits after a context jump — that's free leverage sitting there.
✦
XII ────── Recap
The whole thing in 30 seconds.
Hermes supplies the hands, Qwen 3.8 Max supplies the brain — 2.4 trillion parameters, about 95 billion active at a time, with a million-token memory.
Setup is a folder, a config file and one flag. Then verify with the status command before you trust a single answer.
The trap: a route can advertise the right model name and still be pinned to a retired version, so every request fails. Send one real message to prove it works. And use a provider name your agent knows — invented ones are rejected outright.
Two builds, one prompt each, both worked first time — a 3D jade mechanism and an ink-wash garden, both playable above. It added Chinese characters and a matching seal I never asked for.
The real win isn't the model. It's an agent where the brain is a swappable part, so the next launch costs you two minutes instead of a weekend.
The shortcut
Skip the wiring. Run my exact setup.
Hermes, the Qwen profile, and every other engine — pre-wired inside the Agent OS. 4,000+ founders, 258 documented wins, 38 countries.
The AI Profit Boardroom is where the actual Agent OS lives — the templates, the prompts, the daily rooms, the weekly walkthroughs. Same builds you read about here, taught hands-on inside.
