Qwen 4: Release Date, Leaks, and the Architecture Alibaba Already Shipped

§ 01What we actually know#

§ 02What the Qwen team has said, word for word#

Two sentences carry the whole official record, both from the August 26 release.

From the GitHub repository: “In this release we are opening the weights of Qwen3.8-Flash-Next, a multimodal MoE model that also serves as an early preview of the architecture used in Qwen4.” And: “We are again releasing the architectural changes early, so that the community can examine them before the full Qwen4 model family is built on top of them.”

From the Hugging Face model card: “This experimental preview of the architecture that will underpin Qwen4 is built around a fundamental rethinking of how the core components of modern large language models interact at scale.”

The word “again” is doing work in the first quote. The Qwen team has run this play once before: Qwen3-Next, released in September 2025, previewed the hybrid architecture that the Qwen3.5 family then shipped on, on February 17, 2026. About five months from preview to family. That precedent is the most honest basis anyone has for a Qwen 4 window, and it points at late 2026 or early 2027, not at a conference three weeks away. We will come back to that.

What the team has not said: a month, a size, a lineup, a license, or whether “Qwen 4” will carry the 3.8-Flash-Next configuration up to Max scale or redesign it. Alibaba’s own August 20 earnings release, filed with the SEC, describes the quarter’s model work as “our flagship foundation model Qwen3.8-Max within three months of its prior version, and we opened its model weights with 2.4 trillion parameters.” No Qwen 4.

§ 03The architecture you can already run#

The preview is a real model with a published configuration, and it is the only concrete thing anyone can say about Qwen 4’s design.

The design bet is legible from the numbers: park a large share of the parameters in a lookup table that costs almost nothing to read, keep the per-token compute at 6B, and stretch the context with sparse attention. Whether Qwen 4’s flagship scales that shape to Max size or pairs it with a dense core is the first thing the model card will tell us, and the thing no leak has credibly claimed.

§ 04The rumor chain, dated#

The code is the hardest evidence. Runtimes do not implement an architecture from a rumor. llama.cpp’s pull request 27742, titled “model: add Qwen3.8-Flash-Next (qwen4exp)”, was opened August 26 and merged August 27. Hugging Face Transformers merged “Simplify and fix qwen4 tests” on August 26 and a “[Qwen4 Exp]” mask fix on September 1. vLLM carries open pull requests titled “[Kernel][Qwen] Add qwen4 fuse op” (August 26), “[Qwen4] Support UVA PLE-offload and N-gram parallelism” (August 29), “[Model][Qwen4] Use ragged PLE IDs for CPU offload” (August 30) and “[Qwen4Exp][ROCm] Support FP8 PLE n-gram checkpoints” (September 3). The maintainers named the architecture what Qwen named it. None of that is a Qwen 4 model; all of it is the ground a Qwen 4 model would land on already being paved.

The phantom Coder is the cleanest debunk of the summer. A “Qwen 4 Coder 32B-A3B” entry, complete with a June release date, an Apache 2.0 license and an 82% SWE-bench Verified score, circulated through model databases and the comparison pages that scrape them. No such model exists on the Qwen organizations on Hugging Face, ModelScope or GitHub. The score would have beaten every open-weight coding model on record by a wide margin. The database that carried it, LLMCheck, says it ingested the entry during a July data incident and removed it in an August reconciliation against official sources. If a chart shows “Qwen 4 Coder” beating anything, that is its provenance.

Ox Alpha was never Qwen. The anonymous OpenRouter model that appeared August 20 drew a Qwen guess from some observers. We traced its fingerprints to Z.ai’s GLM family at the time, and Z.ai revealed it as GLM-5.3-Flash on August 26. The full account is on our Ox Alpha page. No anonymous checkpoint on Arena or OpenRouter has been credibly attributed to Qwen 4 since.

§ 05Why Apsara keeps coming up, and why it might not be it#

The September rumor rests on one real pattern. Alibaba Cloud’s Apsara Conference is where Qwen2.5 was revealed on September 19, 2024, and where Qwen3-Max launched on September 24, 2025. This year’s conference is September 22 to 24 at the Hangzhou International Expo Center, dates Alibaba Cloud posted on August 30. A commentator post on X on July 7 put Qwen 4 there, and the idea has circulated since.

Two things cut against it. First, Apsara is not where Qwen generations launch; it is where Qwen ships whatever is ready in late September. Qwen3, the last new generation, launched April 29, 2025, seven months before its conference. Qwen3.8-Max launched August 3, seven weeks before this one. Second, the team’s own precedent for an architecture preview says the family follows in months, not weeks: Qwen3-Next in September 2025, the Qwen3.5 family on February 17, 2026. Applied to an August 26 preview, that arithmetic lands in early 2027.

Our read, labeled as our read: Apsara is the most likely place to hear a Qwen 4 date, and a less likely place to download a Qwen 4 model. A Qwen3.8 refresh on the conference stage, with a Qwen 4 timeline in the keynote, fits every pattern above. A full Qwen 4 family on September 22 would be the fastest preview-to-family turn the team has ever done. We would be glad to be wrong; the page flips either way.

§ 06What shipped instead, and what it costs#

The reference point for any Qwen 4 claim is the generation Alibaba is actually shipping, and August was the busiest month the Qwen team has had.

Six releases in thirty days, two of them open at flagship scale. That cadence is the second reason to read the Qwen 4 rumor calmly: a team shipping this fast does not need Qwen 4 in September to have news at Apsara. It also explains the “within three months of its prior version” line in Alibaba’s earnings release. For how the current flagship compares with its closest open rival, see GLM-5.3 vs Qwen3.8-Max.

We do not route to Qwen today. CellCog’s tiers run on Claude Fable 5.1, Claude Opus 5 and Gemini 3.8 Flash, and we moved to each on the day it shipped. Qwen matters to us for a different reason, and it is the reason this page exists: the Qwen 4 architecture preview is open, Qwen3.8 opened a Max-class model for the first time, and if the Qwen 4 flagship follows, it becomes the first frontier-class model an AI employee platform could credibly run without a vendor API. That is a decision about the engine underneath every employee we operate, and we would rather have the facts dated before the day comes.

§ 07What we are watching for#

We track open-weight frontier models as an operator weighing engines, not as spectators.

• A date from Alibaba. The Apsara keynote on September 22 is the first scheduled chance. A Qwen team post, an Alibaba Cloud release or a model-card commit would count the same.
• A Qwen 4 model card. The moment a Qwen/Qwen4-* repository appears on Hugging Face or ModelScope, or QwenCloud’s release notes list a Qwen 4 identifier, this stops being a rumor page.
• The license line. Qwen3.8-Flash-Next shipped under the Qwen Community License. Whether the Qwen 4 flagship is open, open with restrictions, or API-only is the single fact that changes what this model means for self-hosting.
• The shape at scale. Does the flagship keep the 125B/6B-with-n-gram-table shape and grow it, or pair the sparse-attention stack with a dense Max-class core? The preview card cannot answer that; the first Qwen 4 card will.
• Independent replication. Every Flash-Next number is the Qwen team’s. Third-party runs on the preview, and later on Qwen 4, are what make a real comparison possible.
• Our own routing. If a Qwen 4 model is real, open and better on our axes, we test it on our infrastructure and this page records the result. If it is not, this page records that too.

§ 08The tracker#

This page updates when facts change, not when rumors get louder. As of September 4, 2026: the Qwen 4 architecture is public and running in every major inference engine; the Qwen 4 model is not announced, dated, sized or licensed. The most credible window is not the September conference but the months after it, on the team’s own preview-to-family precedent. When Alibaba moves, the confirmed facts and a graded scorecard of the claims above land here the same day.

§ 09Update log#

This is a living page; when the story moves, the update lands here.

As of September 4, 2026: page opened. No official movement to log.