Alibaba's Qwen3.8-Max spent 16 days writing a tool by itself, and the weights go public next week
Alibaba published full details and benchmarks for Qwen3.8-Max, a 2.4 trillion parameter model built for tasks that run for days. Five case studies cover autonomous coding, paper reproduction, chip design and a simulated year of online retail.
Alibaba has laid out the full picture for Qwen3.8-Max, the flagship model it previewed in mid July, and this time the numbers and the case studies come with it. The model has 2.4 trillion total parameters with 95 billion active per query, and the team says the weights go up on Hugging Face and ModelScope next week. That would make it the first model in the Qwen-Max class you can download and run yourself.
The pitch is not better answers to single questions. It is what the field calls long horizon work, meaning tasks that run for days across hundreds of steps. Alibaba published five case studies to make the point. In the first, the model spent 16 days building a command line tool called oh-my-cli: it turned incoming user requests into GitHub issues, assigned them to itself, wrote the code, ran the tests and iterated. By 30 July it had logged 265 commits, 127 pull requests and 151 issues with no human touching the keyboard. In another, given only a research paper and no starter code, it reproduced all six of the paper’s results over roughly 125 hours of compute, then tested 18 ideas of its own and beat the original method on a maths benchmark by 2.7 points.
Two more runs test planning rather than coding. Asked to design a cryptographic circuit, the model started from a working but bloated design using 8,298 logic gates, the basic switching elements on a chip, and got it down to 678 over about 500 iterations. After an automated layout pass the physical area shrank by 81 percent. In a simulation of a full fiscal year of online retail, built on anonymised Taobao and Tmall data, it started with 100,000 yuan, negotiated with suppliers in plain language, handled returns and typhoons, spotted scammers hidden in the supplier pool, and finished with 416,252 yuan.
What is behind this. Alibaba credits the jump to how it ran reinforcement learning, the training stage where a model is rewarded for good outcomes rather than shown correct answers. Instead of practising on single tasks, it practised in roughly 4,000 different environments covering multi day workflows, nested project folders and several different agent frameworks. Its internal score index across ten plus benchmarks rose from 0.474 to 0.725, then dipped slightly past that point. Worth keeping expectations grounded: all of these numbers are self reported from internal runs, no one outside the company has verified them, and a model that quadruples play money in a simulation has not run a real shop. The competitive backdrop matters too. Moonshot’s Kimi K3 opened its weights on 27 July, and independent testing brought its claims down to earth.
What this means for you: nothing today, unless you rent GPU time by the hour. A 2.4 trillion parameter model is not something you run on a laptop, so in practice most people will meet it through a hosted service. What is genuinely useful is the compatibility choice: Qwen3.8-Max speaks both OpenAI’s and Anthropic’s API formats, so it drops into Claude Code, Codex and similar tools by changing a setting. If you already pay per token for an agent that runs long jobs, that is a real price lever worth testing once the weights land. If you do not, the signal to take away is that the open and closed camps are now roughly one step apart rather than one generation.
Sources
Source: https://qwen.ai/blog?id=qwen3.8
Two research teams solved the same open problem with the same AI model, three hours apart
An MIT doctoral student and two professors independently cracked the same question in quantum cryptography using GPT-5.6 Sol Ultra, and posted their papers to arXiv within three hours of each other. It raises an awkward question about what independent discovery means now.
