Qwen 3.8 Max Just Changed AI Agents Forever

**Julian Goldie** (0:00)
And Qwen 3.8 Max just officially dropped in China, did something that nobody thought was possible two years ago. This model worked completely alone for over 10 days straight, with no human touch in it. It built an entire software project from an empty folder, tested its own work, fixed its own mistakes, and shipped 260 people in the world, ahead of Google, ahead of Meta, ahead of OpenAI, on some of the benchmarks it's free to use right now. And next week, they're giving the weights away. We'll be open source. I've been testing this model for weeks in preview mode. We ran it through 45 different tasks on Goldie bench against Claude Fabel, GPT 5.6 Sol, and Kimi K3. 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 Qwen 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 open weights go public, along with a smaller Qwen 3.8 27b. They will also be going open, and that will 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 open weights 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 and 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 a 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 glue 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 the 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, and you come back to a finished result. And here's the proof today is that Qwen 3.8 Max was built to do this. Alibaba published the 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 us 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 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 that 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 0.6 to 0.853.