Alibaba’s AI coded for 16 days straight and every commit is on GitHub
The New Stack Amanda Caswell ● Covered by 3 sources
Alibaba's new Qwen3.8-Max spent 16 days autonomously coding an app, with every commit posted publicly on GitHub. It's a massive 2.4-trillion-parameter model, but weights drop next week so anyone can check the work.
Alibaba just put its money where its mouth is. Instead of just claiming Qwen3.8-Max can handle long, messy coding jobs, the company let it run wild for 16 days building a command-line tool called oh-my-cli, then published the entire 265-commit, 127-pull-request trail on GitHub under qwen-code-dev-bot/oh-my-cli. You don't have to trust Alibaba's benchmark slides. You can just go read the commits.
The model itself is a beast: 2.4 trillion total parameters, though thanks to a sparse mixture-of-experts design built on the Qwen3.5 architecture, only about 95 billion get activated per token. That's the trick that makes a model this size usable at all, but it doesn't make it cheap to host. Alibaba is pricing API access at $2 per million input tokens and $6 per million output tokens through Model Studio and QwenCloud, and says it'll release the weights on Hugging Face and ModelScope next week — the first Qwen-Max model to go open-weight. Good luck running it yourself, though. Storing and distributing those full weights across high-memory GPU nodes is realistically a job for large enterprises and inference providers, not solo developers with a spare A100.
The timing is not an accident. Moonshot AI's Kimi K3, a 2.8-trillion-parameter model in which Alibaba holds a 36% stake, just topped Arena's frontend coding leaderboard and got so much demand that Moonshot had to pause new subscriptions within 48 hours. Qwen3.8-Max is entering that same lane, and Alibaba clearly wants proof-of-work, not just leaderboard bragging rights. Beyond the 16-day CLI build, the company says the model spent five days reproducing results from a machine learning paper — writing roughly 7,600 lines of code, kicking off 33 GPU training jobs, and even beating the paper's reported AIME24 score by 2.7 points. In a separate chip-design test, it iterated about 500 times to shrink a cryptographic circuit from 8,298 logic gates down to 678, cutting physical area by 81% once run through the open-source OpenROAD tool.
What's actually interesting here isn't the parameter count, it's the harness — the scaffolding around the model that decides whether an agent survives 16 days or dies after 16 minutes. Alibaba seems to understand this: it trained Qwen3.8-Max on realistic multi-day projects and published guides for hooking it into Claude Code, Codex, OpenClaw, Hermes and its own QwenWork, with Model Studio supporting both OpenAI and Anthropic API formats. That's the same lesson Anthropic paid for when it acquired the CI/CD startup Mendral, and the same one GoDaddy learned the hard way when it opened its domain registrar to autonomous agents.
None of this settles the benchmark arms race. Alibaba claims 86.6 on Terminal-Bench 2.1 against GPT-5.6 Sol's 88.8, fifth place in Text Arena, second in Vision Arena, and wins in six of 31 internal comparison tests — with Claude Fable 5 or GPT-5.6 Sol ahead almost everywhere else. Independent labs tend to reproduce these numbers differently anyway. The real test comes once the weights are out in the wild next week, when developers get to see if Qwen3.8-Max still performs once it's off Alibaba's home turf and running on someone else's infrastructure.
My take
The public GitHub trail is the smartest thing Alibaba did here — anyone claiming multi-day autonomous coding without receipts deserves to be ignored, and this is the rare lab that showed its receipts. That said, a 2.4-trillion-parameter model that only a handful of cloud giants can actually self-host isn't really 'open' in any way that matters to the average developer; it's open-washing for enterprises with spare GPU clusters. Watch what happens once independent labs get their hands on the weights next week — that's when we'll know if this is a genuine leap or another leaderboard flex.
Read more about this at: The New Stack
