Alibaba unveils Qwen 3.8 Max, a 2.4-trillion-parameter MoE model with a 16-day autonomy claim

Alibaba says the model can run long, autonomous tasks for more than 16 days without human intervention. A promotional video shows it spending 12 hours independently designing a computer chip from start to finish. Missing from Alibaba’s announcement is a figure that actually matters for judging a sparse MoE model this size: how many of the 2.4 trillion parameters activate per token. That number drives real inference cost, and Alibaba simply hasn’t shared it. The chip-design run is worth watching directly — it shows how the model handled a long, structured task without a human resetting it midway.

A twelve-hour demo proves the model can hold a task together without collapsing partway through. It doesn’t show how often a real deployment would need someone to step in and correct a wrong turn, and a single curated video can’t answer that on its own. The 16-day autonomy figure invites the same skepticism at a larger scale, since staying useful that long depends heavily on how forgiving the underlying task is, not just on the model running it.

Qwen 3.8 Max runs at $2 per million input tokens and $6 per million output tokens, a price aggressive enough to undercut several rival frontier models outright. Alibaba is also opening the weights for Qwen 3.8 Max and a smaller Qwen 3.8-27B variant next week. For a developer picking between models day to day, that price tag will likely carry more weight than any benchmark chart.

Alibaba’s own internal testing puts Qwen 3.8 Max ahead of ChatGPT-5.6 Sol, Claude Opus 4.8, and Claude Fable on several benchmarks, and first place on the Frontend Code Arena leaderboard. Nobody outside Alibaba has reproduced those results yet, so treat them as Alibaba’s own scorecard for now.

The playbook looks familiar. Moonshot AI ran the same combination in July with Kimi K3: a huge headline parameter count paired with pricing built to undercut Western labs, alongside far less architectural disclosure than US companies typically publish.