An AI coded alone for 16 days and built its own tools to do the job

Alibaba says its new model worked on a software project for more than sixteen consecutive days without a human intervening — and that the first thing it did was build the tools it would need to do the work.

The Claim

Qwen3.8-Max, Alibaba's new flagship, is a mixture-of-experts model with 2.4 trillion total parameters, of which roughly 95 billion activate during any given inference. It carries a one-million-token context window.

The headline demonstration is the autonomy run. According to Alibaba, the model worked continuously for over 16 days on a software engineering project, beginning from an empty repository and producing an open-sourced framework called oh-my-cli. What makes the claim notable is not the duration alone but what the model built during it: its own coding harness, with a state machine, a dispatcher and a testing loop — the scaffolding that would normally be supplied to an AI agent by the engineers deploying it.

In other words, before writing the software it had been asked to write, the model constructed the apparatus it needed to keep itself working coherently across a span of days. It committed code to GitHub throughout.

Why Duration Is The Hard Part

Long-horizon autonomy has been the persistent weakness of AI coding systems. Models are strong at bounded tasks — write this function, fix this bug, refactor this module — and degrade over extended sequences. Context is lost, earlier decisions are contradicted, errors compound silently, and the system drifts from the original objective without noticing.

Sixteen days of coherent work on a single project, if it holds up to outside examination, is a substantial jump rather than an incremental one. It is the difference between an assistant that helps a developer and a system that completes a project.

The Sceptical Read

That last conditional matters, and the reaction has not been uniformly admiring.

The performance figures are vendor-reported. Alibaba conducted the run, defined the task, and published the result. There is no independent replication of a 16-day autonomous session, and the nature of the demonstration makes verification difficult — an outside party would need to run an equivalent project under equivalent conditions to test the claim.

The New Stack characterised the release as ‘an API business model wearing an open source jacket,’ a pointed reference to the gap between open-weight positioning and the commercial reality of how the model is actually accessed and monetised.

The industry also has recent reason for caution about vendor benchmark claims. Meta's former chief AI scientist, Yann LeCun, confirmed earlier this year that Llama 4's headline results had been assembled by training multiple checkpoints and selecting the best score on each individual benchmark — publishing a composite no single model had achieved.

The Competitive Picture

Qwen3.8-Max continues a pattern that has defined this quarter. Chinese laboratories briefly held six of the top ten positions on independent model leaderboards in mid-July, including all of the top five. Moonshot AI released Kimi K3's full weights for free download on July 27 — at 2.8 trillion parameters, the largest open-weight model ever published — and OpenAI cut prices on its lower-cost models by up to 80 per cent three days later.

Alibaba's entry adds a distinct claim to that competition. Kimi K3 competed on scale and price. Qwen3.8-Max is competing on autonomy — the ability to be left alone with a task for weeks. If the demonstration is representative rather than a carefully staged best case, it is the more commercially consequential capability of the two.

That is the question outside researchers will now try to answer, and the model being available is what makes answering it possible.