Qwen4 appears in Alibaba’s code, with a million-token context default

Alibaba’s Qwen Code repository is preparing for Qwen4. A change merged on October 28 gives names beginning with qwen4 a default context budget of 1,000,000 tokens when the model catalog has no more specific entry. On October 31, Mark Kretschmann (@mark_k) pointed readers to the change.

That number tells the coding client how much material to budget for a conversation. It does not demonstrate that a model can reliably use a million tokens. The pull request calls the figure reviewer-reported and unconfirmed; its tests used a fake server and made no calls to an actual Qwen4 endpoint. The model catalog can still override the default.

There is another preparatory detail in the provider documentation: a reasoning_effort setting for the Qwen4 family, with xhigh at the top. This describes how the client sends a request, not a measured improvement in reasoning. Together, the changes show software being readied for a new generation before its public specifications are settled.

Lumina (@LuminaBench) also reported an arrival “next week,” citing staff remarks at Apsara. That timetable remains a reported claim: we have not inspected a recording or an official release schedule. The concrete development is readiness in the client code, not a Qwen4 launch.