Sixteen Days, 265 Commits, Zero Humans: Alibaba's New Qwen

Alibaba released a new AI model today, and to show off what it can do, they let it run a software project by itself for 16 days. No human touched it. By the end, the project had 265 commits, 127 pull requests, and 151 issues, and the whole work log is public on GitHub for anyone to audit.

That is the opening number for Qwen3.8-Max, Alibaba's biggest model ever, released this morning. Here's what we actually know, with the sources, because some of this is Alibaba grading its own homework.

What Alibaba shipped

Per Alibaba's announcement blog and press release, Qwen3.8-Max is a 2.4-trillion-parameter mixture-of-experts model with about 95 billion parameters active per query, a 1-million-token context window, and text, image, and video input. It is live on the QwenCloud API today, and the weights go public on Hugging Face and ModelScope next week, which would make it the first open-weight model in Alibaba's top "Max" class. A smaller 27B checkpoint is going open at the same time.

On the numbers Alibaba published, the model broadly matches and sometimes beats Anthropic's Claude Fable 5 on coding and agentic tests. Take that framing with the usual grain of salt, since they are internal evals. The independent check is Arena, the crowdsourced leaderboard: per The Verge's reporting and Alibaba's press release, Qwen3.8-Max ranks fifth in Text Arena, behind only Fable 5 and three models from Anthropic's Opus family, and second in Vision Arena. In frontend coding it trails two Opus models and Moonshot's Kimi K3. It has been out for less than a day, so nobody can honestly call the overall pecking order yet. But fifth-best in the world on a public leaderboard, from a model whose weights will be free next week, is a real result.

The part I keep thinking about

The benchmark table is not the interesting thing in this release. The interesting thing is what Alibaba made the model DO.

Per the company's write-up, they handed it a recent research paper and told it to reproduce the experiments from scratch, no starter code. It spent about 37 hours rebuilding the full pipeline and confirmed the paper's six main findings.

Then it kept going for another 88 hours, running its own improvement loop. Form a hypothesis, write the code, run it on GPUs, analyze, repeat. Eighteen ideas across four rounds, ending with a method that beat the paper's own approach by 2.7 points on a competition math benchmark. Total: roughly five days, 7,600 lines of code, over 1,100 actions, zero humans.

Then they entered it in an online contest against 526 human teams, under a 24-hour limit. It placed ahead of 458 of them.

And the 16-day run from the lede was a real software project with real community feature requests coming in through GitHub issues, which the model claimed, coded, tested, and merged on its own.

Again, this is Alibaba's own demo reel. But the receipts are public repos and competition scoreboards, not just a chart, and that is a different kind of claim than "we scored well on a test."

What it costs

This is where it gets practical for people like me. Per QwenCloud's model card, the API price is $2 per million input tokens and $6 per million output, with cached input at $0.25. One detail that matters more than it sounds: per eesel AI's pricing breakdown, that rate is flat across the entire 1M context window. Most long-context models jack up the per-token price once your input crosses a threshold, which is exactly where long agent sessions quietly get expensive. No tier here.

For comparison, per the same breakdown, Kimi K3 runs about $3 in and $15 out. So Alibaba is pricing its flagship at roughly half of Moonshot's, and far below Anthropic's top tier.

Regular readers know I run Kimi K3 inside Hermes as the daily brain of my agent setup, so the Chinese labs' pricing war is not abstract to me. One more factual detail from Alibaba's blog: they list Hermes, the exact agent framework I run, as one of the harnesses they trained and tested this model on, and per The Decoder the API speaks Anthropic's protocol, so it drops into Claude Code-style tooling with a config change. Whoever this model was built for, it was built for setups like mine.

The catch, and it is a familiar one

Last week I wrote about Kimi K3's open weights: free to download is not free to run. Same story here, bigger. Kimi K3 is 2.8 trillion parameters. Qwen3.8-Max is 2.4 trillion. Per MarkTechPost's analysis, a checkpoint that size is a multi-node datacenter job, not something that fits on anything in my apartment, and the 27B version is the realistic self-hosting path. The weights being public matters for researchers, for fine-tuning shops, and for keeping the closed labs honest on price. It does not mean you or I are running this thing locally next week.

What it does mean is there is now a second Chinese flagship, weeks after Kimi K3, sitting near the top of the public leaderboards at a fraction of American API prices, with the weights about to be free. ByteDance and MiniMax shipped new video models on Friday, per The Verge. The releases are coming weeks apart now, and each one is a little closer to the top.

Next week the weights drop and the whole internet gets to check Alibaba's homework. I will be watching what people find.