Alibaba built a 2.4 trillion parameter model — and they're giving the weights away.
Qwen 3.8 Max just landed, and it is the biggest model anyone has ever promised to hand over for free.
Two point four trillion parameters, a million tokens of memory, and Alibaba say it can code on its own for more than ten days straight.
Next week they release the weights — the first time a Max-class model has ever been opened up like this.
That means the most powerful tier stops being something you rent and starts being something you can own.
Today I'll show you exactly what shipped, what the builders are saying about it, and the one detail in the announcement that matters far more than the headline number.
Stick with me to the end for that part. Let's get into it.
Here's the launch in one breath.
Qwen 3.8 Max is Alibaba's most capable model to date — 2.4 trillion parameters in a sparse mixture-of-experts design, with roughly 95 billion active on any given token.
A one-million-token context window.
It reads text, images, video and documents.
It's live now on QwenCloud at two dollars in and six dollars out per million tokens.
And next week the weights go public — along with a smaller Qwen3.8-27B you can run yourself.
2.4Ttotal parameters
95Bactive per token
1Mtoken context window
2/2builds worked first try (mine)
II ────── What actually shipped
The headline is the size. The story is the licence.
Big models arrive most weeks now. This one is different for a reason that has nothing to do with the parameter count.
Alibaba announced it themselves, and the second line is the one to read twice:
Straight from Qwen →
📢 Meet Qwen3.8-Max — our most capable model to date. Next week, the open weights of Qwen3.8-Max will be released, and Qwen3.8-27B is also going open-weights to meet you all! 🎉 Qwen3.8-Max, a new bar for coding and cowork at 2.4T parameters — autonomous coding: 10+ days…
— Qwen (@Alibaba_Qwen)
Three things in there matter:
The weights are coming. This is the first time Alibaba has opened a Max-class model. Their top tier, not a stripped-down sibling.
There's a small one too. Qwen3.8-27B goes open at the same time — that's the size that actually runs on a decent machine at home.
They're claiming ten-day autonomous coding runs. Not ten-minute tasks. Multi-day agent work without a human resetting it.
The builder community caught the significance immediately:
And the reaction →
Qwen3.8-Max is going open-weight next week. It marks the first time Qwen will open-source the weights of a Qwen-Max-class model. The second >2T open-weight model after Kimi K3, the benchmark results look amazing, hopefully it's MIT license. Let's go oss LLMs!
— Yuchen Jin (@Yuchenj_UW)
Frontier-scale models have been rentals. This one is about to become property.
"Can I actually run a 2.4 trillion parameter model?"
Almost certainly not at home — a model that size needs serious server hardware even with only 95 billion parameters active at a time.
But that isn't the point of open weights. It means anyone can host it, inspect it, fine-tune it and build on it without asking permission.
And the 27B released alongside it is the one for your own machine.
✦
III ────── The claim to watch
They say it's second only to Fable 5. They published no table.
Here's the part I want you to hold lightly, because everyone else is repeating it as fact.
Alibaba positioned Qwen 3.8 Max as trailing only Anthropic's Claude Fable 5 — the strongest claim any lab has made against the current leader.
But as of writing there is no published benchmark table, no model card, and no independent testing behind that ranking.
That doesn't make it false. Big claims from big labs often hold up. It just means nobody has checked yet.
What to do with a claim like that: treat it as a hypothesis, not a result. The weights land next week — and the moment they do, independent people will test it properly and we'll all know.
That week-long gap between "we say it's second best" and "here's the evidence" is where most AI hype lives. You can simply wait it out.
Open weights close this gap permanently — once anyone can run it, marketing stops deciding the ranking.
"So is it actually better than the model I'm using?"
Nobody knows yet, and anyone telling you otherwise this week is guessing.
What we do know is verifiable: the size, the context window, the price, and the fact that the weights are being released.
Judge it on your own work when you can run it — that's the only benchmark that decides your afternoon.
✦
IV ────── The number nobody's talking about
Ten days of coding without a human.
The parameter count got the headlines. This is the line that actually changes how you'd use it.
Alibaba describe Qwen 3.8 Max as capable of ten-plus days of autonomous coding, and pitch it as "a new bar for coding and cowork".
Think about what that implies. Not a model that answers well — a model built to stay coherent across a task longer than most people's holidays.
That's the difference between an assistant and an employee.
An assistant needs you to come back every few minutes.
Something that holds a job for ten days needs a brief on Monday and a review on Friday.
It's also exactly what a million-token context window is for. Enormous memory only matters if the model can use it over a long horizon without losing the plot.
Whether it delivers on that is the thing I most want to test — and it's testable, which is more than can be said for the ranking claim.
"What would I even give an agent for ten days?"
A migration. A test suite. A whole content backlog. A refactor you've been avoiding for a year.
The jobs you never start because they don't fit in one sitting are precisely the jobs a long-horizon agent exists for.
Most people underuse these tools because they only ever hand them ten-minute problems.
✦
V ────── I ran it myself
So I stopped reading about it and gave it two jobs.
Everything above this point is what Alibaba announced. Here's the part nobody else on your feed has: what it actually does when you hand it work.
I wired Qwen 3.8 Max into my agent as a profile — one config file, about two minutes — and gave it two builds. One prompt each. No follow-ups, no corrections, no fixing anything by hand afterwards.
Both worked on the first attempt. Complete files, zero console errors, animating the moment they loaded. I checked them in a real browser rather than taking the model's word for it.
The Ink Garden is the one that surprised me. I asked for ink-wash branches and a seal stamp — it added Chinese characters for "ink garden" in the corner and matched them inside the seal, which I never requested. That's a model with genuine cultural range, not a translation layer bolted onto a Western model.
The Jade Engine is the harder test: real 3D, interlocking geometry, a HUD number that has to move. It nailed the brief including the live RPM counter.
The honest caveat: two builds is not a benchmark. It tells you the model is competent and follows a brief closely — not where it sits against Fable 5. That ranking still needs proper testing, which gets much easier the moment the weights are public.
"How do I point my own agent at it?"
If you run Hermes, it's a profile — a small config file naming the model and the provider, then hermes -p qwen-3-8. Any agent that lets you set a custom model and endpoint works the same way.
Ten minutes at most, and you can A/B it against whatever you use now on real work.
✦
VI ────── My story · why this matters
I was you. Then I stopped renting my brain.
Before
Everything I built sat on top of somebody else's model.
A price change, a rate limit or a deprecation notice could reshape my week without warning.
I'd wake up to "this model is being retired" and spend a day rebuilding things that were working perfectly the night before.
Then I built a system where the model is a swappable part — and started keeping open ones in the mix.
After
Now a launch like this is a good day rather than a disruption.
If the weights are strong, I plug them in beside everything else and test on real work.
If they're not, I've lost an afternoon.
Either way nobody can retire, reprice or restrict the models I actually own.
You can build this too. The system first, then the models come to you.
VII ────── The receipts
Real people. Real builds. Inside the Boardroom right now.
The members running open and closed models side by side — 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.
The weights land in days. Everyone will have an opinion; almost nobody will have run it.
So here's the deal.
When it drops, give it one real job from your own week — not a quiz, not a riddle. Something you'd otherwise do yourself.
Then judge it on what comes back.
The people who form views from threads are always one step behind the people who tested it on Tuesday.
Be one of those people. One real job. Next week.
✦
VIII ────── Old way vs new way
Most people pick a side. Operators keep both.
The old way: pick a lab, build everything on it, defend the choice.
When it gets more expensive or gets retired, absorb the damage and rebuild.
Open versus closed becomes an argument you have online instead of a decision you make per job.
The new way: the system is permanent, models are parts.
Closed frontier models for the hardest work while they lead.
Open weights for anything you need to own, run privately, or protect from someone else's pricing page.
A launch like this doesn't threaten your setup — it adds an option to it.
✦ ✦ ✦
Get the whole operating system
Want new models plugged in for you?
Every engine in one dashboard, swappable, tested on real work — that's the Agent OS. Join the AI Profit Boardroom and you get:
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
Model-routing playbooks — which engine for which job, kept current as the field moves
Agent Kanban — Planner → Builder → Reviewer teams that ship while you sleep
The memory vault — an Obsidian brain your agents actually read
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 every launch 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.
"Open models are always behind the closed ones."
That was reliably true and it's stopped being reliable.
This is a 2.4-trillion-parameter Max-class model being opened, following another two-trillion-plus open release before it.
The gap is now measured in months, sometimes weeks — and for a lot of everyday work it closed a while ago.
"There's a new model every week, so why care about this one?"
Most launches change which rental is cheapest. This one changes what you can own.
An open Max-class model can be run privately, fine-tuned on your own data, and kept running long after the lab moves on.
That's a different category of event to a price cut.
"I'll wait for someone to tell me if it's good."
Everyone is waiting for that, which is exactly why the people who test it first get a head start measured in weeks.
Testing it costs you one afternoon and one real task.
Reading forty threads about it costs more than that and teaches you less.
✦
XI ────── The SOP
How to meet a launch like this — five steps.
1 · Separate the verifiable from the claimed. Size, context window, price and licence are facts. "Second only to X" is marketing until someone publishes a table. Write the two lists separately and you'll never get carried away.
2 · Wait for the weights before you rank it. Open weights mean independent testing within days. A week of patience turns every argument into a measurement.
3 · Grab the small one too. The 27B is the version that runs on your own machine. For most people that release matters more day to day than the 2.4T flagship.
4 · Test it on a real job, not a riddle. Give it something from your actual week. Puzzle performance tells you nothing about whether it'll finish your migration.
5 · Keep it beside your others, not instead of them. Add it as another option in your setup and let real work decide. The winner should be chosen by results, not by whichever launch you read most recently.
✦
XII ────── Recap
The whole launch in 30 seconds.
Qwen 3.8 Max is Alibaba's most capable model to date — 2.4 trillion parameters, about 95 billion active per token, a 1M-token context window, multimodal, live on QwenCloud at $2 in / $6 out per million.
The weights go public next week — the first time a Max-class Qwen has been opened — along with a Qwen3.8-27B you can actually run yourself.
They claim more than ten days of autonomous coding and "a new bar for coding and cowork". That's the claim worth testing, because it's the one that changes how you'd use it.
They also claim it trails only Claude Fable 5 — with no benchmark table, no model card and no independent testing published. Hold that one lightly until the weights land.
I tested it: wired it into my agent as a profile and gave it two one-shot builds — a 3D jade mechanism and an ink-wash garden. Both landed first try with zero errors, and it added Chinese characters and a matching seal I never asked for.
The move: note the verifiable facts, wait days rather than argue, then give it one real job from your own week and judge it on what comes back.
The shortcut
Skip the setup. Run my exact system.
Every engine — open and closed — pre-wired and swappable 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.
