Alibaba's own Qwen Code repo now sizes Qwen 4 at 1M tokens

Alibaba's open-source coding agent, Qwen Code, now has a rule for a model family Alibaba has not released. A pull request merged at 05:26 UTC on October 28 sizes every model id that starts with 'qwen4' at a 1,000,000-token context window, and its tests check that rule against qwen4, qwen4-max, qwen4-plus and qwen4-coder-plus.

WHAT THE CODE SAYS

The rule sits in the tool's context-window table beside the existing Qwen 3.x row. Its comment calls 1M 'the forward default for the qwen4 generation', so later 4.x releases need no further change, and it says the models.dev catalog still wins once it lists a Qwen 4 model with a different window. The author, Shaojin Wen, lists Alibaba as employer on GitHub. The pull request says no qwen4 id exists yet in models.dev or in the tool's built-in catalog.

THE OTHER IDS

A qwen4-omni-120b-think id has sat in the repo's multimodal test fixtures since September 16. Earlier on October 28 the same author also sized invite-only aliases such as 'qwen-latest-series-invite-202610-m4' at 1M, saying the whole series shares that window. The repo does not say which model those invite aliases serve.

WHAT THIS DOES NOT PROVE

The 1M figure is the tool's default, not a specification Alibaba has published, and the same tests use spellings that are plainly placeholders, such as qwen4.5-max and a dated qwen4-max snapshot. A name in a test is not a product. Neither OpenRouter nor models.dev lists any Qwen 4 model as of 12:00 UTC on October 29.

THE CONTEXT

Alibaba said Qwen 4 was in training on September 22. At Alibaba Cloud's Apsara Conference in Amsterdam on October 29, attendee @jtdavies quoted Alibaba Cloud CTO Feifei Li on stage: 'Qwen4 next week'. Alibaba has not posted that itself. The repo change was first spotted on X by @Mr_Salio.

WHAT WE DO NOT KNOW

The release date, the lineup, the real context window and the price.