Qwen4 Code Leak: What the Qwen Code Commit Shows, and What Claims Are Unverified

Reports on November 3 say that a code change in Alibaba's open-source tooling "leaked" unannounced Qwen4 models and a 1 million token context window. We read the actual commit. It is real, and it is smaller than the headlines. It adds one rule that sizes any model id starting with qwen4 at 1M tokens. The names qwen4-coder-plus and the others appear as test inputs. No weights, endpoints or model cards are involved.

This is a claimed-versus-verified post. Qwen4 is the model family many readers are waiting for, and our Qwen4 architecture preview coverage explains why. The date and the lineup are still not confirmed by Alibaba in anything we could read.

TL;DR: claimed vs verified

What the commit actually does

The commit is 9ef52e1 in QwenLM/qwen-code, dated October 28, 2026. Its message explains the motive: "The Qwen 1M rules stopped at qwen3.x, so every qwen4 spelling fell through to the 256K family row and auto-compaction fired at 22% of the real window."

So the change fixes a practical bug. The tool needs to know each model's context window to decide when to compact a long conversation. If it did not recognize a qwen4 id, it would assume 256K and compact far too early. The authors added a family row.

The code comment is the most useful line: "1M is the forward default for the qwen4 generation, so qwen4.x releases need no further code change. The catalog still wins once models.dev publishes a qwen4 entry with a different window."

Read that carefully. It says the window is a guess made in advance that the catalog can override. A test even checks that a catalog entry of 400,000 tokens outranks the family row. That is why the claim "Qwen4 has a 1M window" is not yet a fact about the model.

A ruler extending to one million tokens, the Qwen4 default context window set in a Qwen Code commit

The model names are test strings

The test file the commit changed checks these ids: qwen4, qwen4-max, qwen4-plus, qwen4-coder-plus, qwen4.5-max, qwen4-max-2026-10-01 and openai/qwen4-max. All must return 1,000,000.

A developer writing a regex test lists plausible spellings, including a hypothetical qwen4.5-max and a dated snapshot. Nobody would read qwen4.5-max as an announced model. By the same logic, qwen4-coder-plus is evidence that the Qwen Code team expects a coder variant name, or at least wants to be safe if one appears. It does not prove one exists. Developers who watch these repos know that tool authors often add model ids before launch, and also that they sometimes guess.

A second commit: effort tiers for qwen4

A later commit, 80273bd (October 31), is titled "extend built-in reasoning effort to DeepSeek-V4 and new Qwen prefixes." Its documentation changes say the /effort tiers pass through for qwen3.8-max ids, invitation-only qwen-latest-series-invite-* previews, and the qwen4* family. The ladder stops at xhigh, so a configured max is clamped.

This is stronger evidence that engineers are preparing for a Qwen4 API with tiered reasoning effort, because it describes behavior of the provider, DashScope, and not only a context number. Still, it is client documentation. Alibaba's own model pages remain the source of truth.

What about qwen4-omni-120b-think?

The most striking claim in the reports is a multimodal reasoning model named qwen4-omni-120b-think, said to appear in the repository's test fixtures. We could not verify it. We read the two commits above and searched by commit message. The GitHub code search API needed authentication, so we could not grep the whole repo for the string. The name did not appear in the commits we read.

Until someone links the exact file and line, treat the model name as a rumor. If you see it cited, ask for the path. A claim that cannot be traced to a file is not a leak.

The launch date claim

Posts from the Apsara conference in Amsterdam reportedly quote Alibaba Cloud CTO Feifei Li as saying Qwen4 comes "next week." Counting from the posts of October 29, that points to November 2 to 8. We did not find Alibaba publishing that statement on its own site or accounts, and an earlier Apsara talk on September 22 reportedly said only that Qwen4 was in training with no date. One report mentions that the developer who wrote the 1M rule called the number an inference.

The honest summary: signs point to an imminent release, but a date is unconfirmed. Our Qwen3.8-Flash coverage shows Alibaba has staged releases before, moving from preview to hosted launch to open weights over weeks. A Qwen4 announcement does not have to mean open weights on day one.

A key beside a half-open gate representing the unreleased Qwen4 models that exist only as names in a code repository

Timeline: how the Qwen4 signals built up

Seen in order, the pattern is consistent with a staged launch. The strongest items are Alibaba's own earlier releases, which already ship a 1M default on the Qwen3.8 line. That makes the commit less surprising: the tool authors extended an existing rule from qwen3.x to the next generation.

How strong is code as evidence?

Not all repository evidence is equal. A rough ladder, from weakest to strongest:

1. A string in a test. Shows what a developer typed. Could be a guess.
2. A default in config. Shows what the team plans for. Can be overridden, as the catalog override test here proves.
3. Provider documentation in the tool. Shows the behavior the team expects from the API. Closer to fact, but still second-hand.
4. A model catalog entry or API listing. Shows the model exists and reports its own limits.
5. Weights and a model card. The release itself.

The Qwen Code commits sit at rungs one to three. That is useful, since it tells you Alibaba's own tool team is ready. It is not rung four or five. When a report calls a test string a "leak," ask which rung it is on.

How to check this yourself

1. Open the commit and read tokenLimits.ts and its test. It is a small diff, so it takes minutes.
2. Look on Alibaba's Hugging Face organization and QwenCloud model list for a qwen4 entry. As of this writing we could not confirm one.
3. Watch for the model catalog entry the code comment mentions. When it appears, its context window replaces the 1M guess.
4. Be careful with third-party "release date" blogs. Several we found cite a single social post.

What this means for builders

If you run Qwen models through coding agents, you do not need to change anything yet. When Qwen4 ships, Qwen Code already handles the context sizing and effort tiers for ids that start with qwen4. Other harnesses may not, and the same bug the commit fixes, premature compaction at a quarter of the real window, can hit them. If you use OpenCode with Qwen, check your configured context limit on day one.

For planning, base decisions on what exists: the Qwen3.8-Flash-Next 125B release, the Qwen3.8-Max open weights, and how they compare in our Fable 5, Grok 4.6, GPT-5.6 Sol and Qwen3.8-Max comparison. Hardware owners can see what a Qwen4-architecture preview needs in Qwen3.8-Flash-Next in FP4 on 96 GiB.

Why a small commit still matters

Open-source tooling is where launches surface first. Inference engines, coding agents and gateways must add model ids ahead of time so that day-one users do not hit errors. That is why pull requests keep becoming news before announcements. It is also why readers should separate three things: a name in a test, a default in a config, and a published model card. Only the third is a release.

Here the evidence supports one narrow conclusion: Alibaba's coding-agent team expects a Qwen4 generation with long context and tiered reasoning effort, and wants support in place. It does not yet show the lineup, the sizes, the licence or the date.

We will update this post when Alibaba publishes a model page or weights.

Related reading

• Qwen3.8-Flash and the Qwen4 architecture preview
• OpenCode adds Qwen3.8-Flash 125B, a Qwen4 preview
• Qwen3.8-Flash-Next 125B MoE release
• Qwen3.8-Max open weights on Hugging Face
• Fable 5 vs Grok 4.6 vs GPT-5.6 Sol vs Qwen3.8-Max
• Qwen3.8-Flash-Next in FP4 on 96 GiB

Details reflect the public Qwen Code repository as read on November 3, 2026. Model names and dates may change.