Qwen 3.8 Max Just Changed AI Agents Forever

And 3.8 Max just officially dropped, and China did something that nobody thought was possible two years ago. This model worked completely alone for over 10 days straight with no human touching it. It built an entire software project from an empty folder, tested its own work, fixed its own mistakes, and shipped 262 in the world ahead of Google, ahead of Meta, ahead of OpenAI on some of the benchmarks. It's free to use right now.

Next week, they're giving the weights away. It will be open source. I've been testing this model for weeks in preview mode. We ran it through 45 different tasks on GoldieBench against Claude Fable, G55.6 Sol, and KimiK3.

We've already plugged it into our Agent OS, so this isn't me reading a press release. This is what actually happened, where it wins, where it fails, and where it falls short, plus how to set it up yourself. So first of all, the release. Alibaba just dropped Quen 3.8 Max today.

As of recording this, they released it about 30 minutes ago, August the 3rd. It's their biggest model ever, 2.4 trillion parameters with 95 billion active at any moment, a 1 million token context window. And next week, the OpenWeights go public, along with a smaller Quen 3.8 27b. They'll also be going open, and that'll probably be pretty good for hosting locally and running free local AI as well.

Alibaba has never open sourced a Max class model before. This is the first time any lab, American or Chinese, has released OpenWeights at this scale. However, DeepSeek have just released their latest version. There's tons of new interesting updates coming out of China, and this is getting wild now.

So stop and think about how strange this is. Two years ago, everyone said Chinese AI was years behind. Today, the number one lab in the world is Anthropic, and the number two lab in the world on some of the benchmarks right now is Alibaba. That's not Google.

That's not Meta. And the American labs at the top keep their best models locked behind an API you rent. The Chinese lab is handing you the weights to download and keep forever. That's the complete opposite of what everyone predicted.

But the ranking isn't even the story. The story is what I call the walkaway engine. And here's what I mean. Up until now, using AI has worked one way.

You type a prompt, the AI answers, you check it, you type again, you're glued to the chair, right? You're the engine and the AI is just a tool that you keep pushing. Now, as AI has evolved this year, the walkaway engine flips out. So you can give the AI one goal, you walk away, it runs a loop on its own, does the work, you check the work, and it can even find what's broken, fix it, go again for hours, for days, you come back to a finished result.

And here's the proof today is that when 3.8 Max was built to do this, Alibaba published a proof and you can go check every piece of it yourself. Now, here's the first example. They actually asked the model to build a coding tool from scratch on an empty folder. No starter code.

Then they left it alone. The model set up its own system for taking in feedback, turning feedback into tasks, doing the tasks, testing the results, and merging the finished work. After about 16 days of fully autonomous running, the public GitHub repo showed two commits, 127 pull requests, and 151 issues, all handled by the AI. The whole project trace is public on GitHub right now.

Anyone can actually audit it themselves. Now, think about that loop. It wasn't just writing code. It was collecting requests, deciding what to build next, building it, testing it, catching its own failures, and then repeating them.

That's a worker, not just a standard chatbot. Now, the second example is they handed a fresh research paper and said, rebuild this experiment, then beat it. No starter code, just the paper and some GPUs. The model worked alone for about five days, around 125 hours straight.

It wrote roughly 7,600 lines of code and ran 33 rounds of GPU training. First, it rebuilt the whole experiment and matched the paper. Then it invented 18 of its own improvement ideas, tested them in rounds, and beat the original paper's method by 2.7 points on a hard math benchmark. It didn't copy the research.

It improved the research alone. The third example, and this is probably my favorite, is Alibaba entered it into a real online competition where 528 human teams were competing. The task was reading customer service chats, including screenshots, and figuring out what each customer wanted. The model got 24 hours.

It read the rules, built its whole solution, submitted 45 times, and used the feedback from each attempt to improve the next. Its score climbed from 0.60 to 0.853. It beat 408 of the 526 teams from humans. That's 8% of the field with one AI in one single day.

Now, here's what our own testing showed, because you should never just trust a company's demos, never just trust the benchmarks. We actually put Quen 3.8 max through 45 tasks on GoldieBench. You can check them out at goldiebench.com. Games, websites, apps, physics, graphics, head-to-head against Claude, Fable 5, G55.6 Sol, and Kimi K3.

And it landed in the top 5 of every model we've ever tested. And here's a key part. It's really one of the only open-source models anywhere near that level. And it's Chinese.

Kimi K3 and Quen 3.8 are the first open models in history to genuinely challenge the closed American frontier. Now, some of the games Quen built came out better than what G55.6 Sol actually produced. Again, you can check them for yourself if you don't believe me, or you could test this stuff out yourself and build it yourself. But the thing that I'm saying here is that, like, when it's built now, this isn't like generic AI slop.

It actually pulled in custom designs I didn't see in from any other model before. And now a quick pause here, because if you want to skip the setup entirely, we've actually built Quen 3.8 Max into the agent operating system inside the Aircraft Boarding. We had it running before the official API even dropped, using the Coder CLI plugged straight into Claude code. And now the zip file comes with Quen 3.8 already connected, plus a Quen 3.8 Masterclass to step-by-step install video and daily tutorials on Quen agent workflows as this model keeps evolving.

You also get four coaching calls every week, where you can ask live questions about where and how you should set up Quen inside your business, and over 3,000 business owners with Quen and other models plugged in. Link in the comments description or go to the AIprofitable.com. So let's talk about rankings properly, because I want to keep this honest. On the text arena leaderboard, where real people just vote on millions of AI model answers, Alibaba just hit the number two lab spot in the world with a score of 1,496.

Anthropx Claude Fable 5 sits at number one with 1,509. Meta is third. Google is fourth. Moonshot, another Chinese AI lab behind Kimi, is fifth.

OpenAI is sixth. Now when you think about that list, two Chinese models, two Chinese labs, just placed in the top five, and one of them is 13 points off the world number one. So no, Quen 3.8 Max is not the best model on Earth. Claude still wins most of the hardest coding benchmarks.

And honestly, Claude is still my favorite daily driver. Nobody serious is claiming China took the, you know, took the number one spot today. But look at what Quen actually wins on, right? On OS World Verified, which tests an AI using a real computer, for example, clicking, typing, moving through apps, it scored 86.1.

That beats Claude Opus 4.8 at 83.4, and Claude Fable 5 at 85. On PaperBench, reproducing real research, it scored 93, which is ahead of GPT 5.6. Sol at 90.5. And on the vision side, it's dominant.

Reading documents, charts, screenshots, and long videos better than almost anything out there. It can watch over 100 hours of video and build a searchable memory of everything in it. And unlike Claude, it generates images and videos. Claude can't do that at all.

With Quen, you actually get it for free. Which brings me to the access point, because this is where the Chinese angle gets real for you. You can use Quen 3.8 Max right now at chat.quen.ai, without even logging in. So there's no signup or no subscription option even exists in the chat.

You can code, you can build websites, you can generate images, you can generate videos all for free. And people who run models locally already love this family. Quen 3.6 is one of the most run local models in the world. And next week, you'll be able to run Quen 3.8 on your own hardware too, if you have a good setup.

Now for me personally, on a Mac Studio, probably not going to work so good. But if you have a DGX Spark, if you have a RTX 5090, something like that, it's going to work beautifully and it's free. That is wild when you think about it. Now, every time I cover a Chinese model, the same free comments show up, which is, it's not really free, but you can get access at chat.quen.com.

Other people say, well, Chinese models can't compete with US models. It's number two in the world on human votes and it beat Claude at Computeus. Other people say, well, it's not really open source, is it? But the weights drop next week and they open source every model line eventually.

That's been their pattern for years. Now, here's why this matters strategically. The Chinese labs are competing on openness and price because that's how you win the other side when everyone else has a brand, right? Every time they open a frontier class model, the cost of intelligence falls to everyone on the planet.

You running a small business benefit directly from competition. Frontier AI that cost a fortune to access two years ago is now free in a browser tab because two countries are racing each other. You didn't even have to do anything to earn that. You just have to actually use it.

Now, the honest downsides, because I promise you the truth here. Sometimes when I've been testing it, especially in preview, during peak hours, it can be super slow, you know, especially when half the world is hammering a free frontier model. It's going to happen, right? You can wait a minute or more per response sometimes.

Just bear that in mind, especially if you're on the free plan. The fix is simple. Just run multiple chats in parallel. One coding website, for example.

It could be one making images, could be one building an app. That's how you should use every AI anyway, right? Run them in parallel, have agents orchestrating them for you. We can do that with our agent OS, which is inside the airprofit boardroom.

And also it needs feedback, right? It's not like, for example, Fable 5 where it just nails everything first time round. With Quen, you can go back and forth a bit more, test the output, tell it what's broken, let it fix it. And after two or three rounds, usually you get where you want to be.

And don't use its deep research feature. When I personally tested it, challenged the sources, it actually admitted that part of the reports were fabricated. It hallucinated a benchmark that traced back to an Instagram post. So research stays with other tools and building stays with Quen.

That's the way that would look at it. Now, none of those kill it, right? Free, open, top five in the world is still an incredible deal. You just need to know where it's strong.

And here's the old way versus the new way, laid out plain for you. So the old way is like you pick out one AI company, probably American, pay whatever they charge, use their app. And when they change the rules, you just adjust, right? Your whole business workflow depends on one provider.

When a new model drops, you have to start over from scratch again. That's what most people are doing. The new way is your system and your AI brains are separate. So Quen, even though it's a great model, it's just a brain.

Your agent, open quote code, for example, is a hand. So these are agent harnesses that do stuff for you. Your obsidian vault is the memory and the agent OS is a body that holds it all together. So when a better brain drops, doesn't matter if it's American or Chinese, you can just swap it in and everything else stays the same.

Your automation workflows don't actually break. And here's the detail in this release that proves a whole industry moved this way. So Alibaba trained Quen 3.8 Max to run inside of a company's tools. Their own documents show it working inside Claude Code, inside Codex, inside OpenCore and inside Hermes.

The API literally speaks Anthropx protocol. You point Claude Code at Quen with a couple of settings and it just works. They actually publish the exact OpenCore config in the release notes. So a Chinese lab is shipping instructions for plugging their brain into American tools.

They want you to swap. They want you to test out this brain in different agent harnesses. And that's exactly how I run it too. Quen 3.8 plugged into, for example, Claude Code through the code of CLI or, for example, through OpenRouter sitting inside our agent OS next to Claude.

Hermes, Kimi K3 and GLM 5.2. So you've got one dashboard with every brain and every file it builds saved in one workspace because the memory lives in Obsidian, which is just free markdown files on your own computer and every brain knows your business from day one, which means if you plug in, for example, Quen today, it's going to have full context and view your business and everyone else. If Fable 6 drops tomorrow or Quen 4 drops next month, we just swap it in and nothing else changes. The models change constantly, but the system stays the same.

And that's how you stay current without rebuilding your life after every launch. Now, let me deal with what some of you are thinking. You're thinking, I'm non-technical. This is for developers, right?

But look at what the model actually did in most tests. It wrote all the code. It ran all the tests, right? It fixed all the bugs.

The human's job was just one thing, which is describe your goal clearly. In the 10-day coding run, the humans wrote requirements in plain words and the AI just did everything else. The technical skill you think you might need is exactly the part of the model that this now handles, right? This can automate the technical stuff for you.

In fact, if anything, the tables have turned so that non-technical people can use this better than technical people. Now, the skill you actually need is knowing what to ask for. That's a business skill, not a coding skill. And with tools like, for example, Blender, MCP Connected, through Hermes, Quen can now operate software.

These take months to learn, right? For example, 3D modeling, product renders, promo shots from a plain English sentence. And the second one that a lot of people say is like, I can't trust an AI to work without watching it. And that was true before, but that's why the old way existed.

And if you look at how these new runs work, well, the model tests its own output. In the coding run, for example, every change triggered builds and tests before it emerged. And failures got routed back for fixes automatically. In an actual chip design test Alibaba published, the model ran about 500 turns of work at every version, and every version had to pass automated correctness checks before it counted.

It took a circuit from 8,298 logic gates down to 678. It never got to fake result because the checking was built into the loop. The trust doesn't come from hoping the AI, right? It comes from the loop that verifies everything.

And the third one that some people say is like, I'll wait until the AI settles down. Here's the problem with waiting. 458 human teams just got beaten by this thing in 24 hours. The teams that finished above it were the ones using every AI tool available.

That's the split forming right now. Not humans versus AI, but humans with agents versus humans without them. With China and the US both shipping frontier models this fast, the gap between the two groups grows every single month. The people running these loops are compounding work whilst they sleep.

So what should you actually do this week? First, check it out. You can go to chat.quen.ai today to test it. You don't even need to log in.

You can give it a real task from your business. Build a landing page. It could be generating images. It could be, for example, judging it on your own work, right?

Second, run these things in parallel. So I would actually have like three different tabs open. Website in one, images in another, an app in the third, 10 times the output in the same sitting. And then third, when the weights actually drop this week, that's when the swapping starts.

So Quen 3.8 can go into agent setup, side-by-side with Claude, compared head-to-head on actual tasks, if you'd like to run local models. Now, if you want all of this just done for you, this is exactly what the agent operating system inside the Air Profitable Boarding was built for. One dashboard where you plug in all your agents, your Claude, your Hermes, your OpenCLAW, all your favorite AI agents directly into the system with Quen 3.8 Max already connected and ready the moment you install it. So you get the agent OS as a zip file, the Quen masterclass, a video install tutorial, a 30-day roadmap for putting Quen agents to work, getting you more leads and more customers and daily updates as we improve the system with every release.

Plus you get four weekly coaching calls where you can ask live questions about your Quen and agent OS setup, a prompt library built for agent goals that actually finish, and a member map so you can meet people near your building with these tools, with someone online 24-7 to help. Link in the comments description or go to theairprofitboard.com to get access. Here's the bottom line. Quen 3.8 Max is the first open weights model built from the ground up to be a walk-away engine.

You can set the goal. You can leave. You can come back to finish work. It just put a Chinese lab at number two in the world.

The weights go public next week. It's free in your browser today and it plugs straight into the tools you already have. The question isn't whether China caught up anymore. That's been answered.

The question is what you're going to hand this thing first. I'll see you in the next one. Cheers, bye-bye.