Qwen4 models will get 1M-token context windows, @ItsmeAjayKV says, with Qwen4-27b as the one exception. The post links to a commit in the Qwen Code repository. The commit sizes Qwen4 IDs at 1M in Qwen Code's detector, but it does not mention the 27B exception.

Image shared by @ItsmeAjayKV on X
Image shared by @ItsmeAjayKV on X

Image credit: @ItsmeAjayKV on X

The commit is 3a39309, part of PR #15762, dated October 28, 2026. It adds one pattern rule, ^qwen4, that sizes any qwen4-named model at 1,000,000 tokens in Qwen Code's context-window detector. According to the PR description, those IDs previously fell through to the 256K Qwen family default, so auto-compaction fired at about 22% of the real window. The commit changes only the context window; it does not cover output limits.

What the commit does and does not do

The rule is a client-side default. A model-catalog entry for a specific Qwen4 model overrides it. The tests cover bare qwen4, qwen4-max, qwen4-plus, qwen4-coder-plus, a dotted minor (qwen4.5-max), a dated snapshot and a provider-prefixed ID, all expected at 1,000,000. The PR says later minor releases should need no further code change.

The 27B exception

The post says Qwen4-27b is the exception. The commit does not mention a 27B model, and neither Alibaba nor the Qwen team has confirmed a carve-out.

Where Qwen4 stands

Secondary coverage of Alibaba's Apsara 2026 event reported that Qwen 4 was still in training, with no date, size or licence given. A third-party tracker said in late October that no Qwen 4 variant had shipped. Qwen3.8-Max-0902 already launched with a 1M-token window, as our earlier report covers, so a 1M figure for the next generation is plausible.

Our read: treat 1M for Qwen4 as the Qwen Code team's planning default until Alibaba publishes a model card with its own context figure.

Who should care

If you use Qwen models in Qwen Code, this commit changes when compaction fires once Qwen4 IDs appear. Check output limits separately, since this commit leaves them alone. For the wider Qwen4 picture, see our report on the lineup.

Sources