Every few weeks, an AI lab shows a slide and the internet decides the product already exists. Qwen4 is the latest case. Search for it and you will find pages that read like launch coverage. They are not. Nothing has shipped.
I spent the last couple of days going through Alibaba's conference material, the third-party trackers and the code repositories people are quoting. My conclusion is simple. The four names are real. Almost everything else you are reading is inference, and some of it is dressed up as fact.
What Alibaba has actually shown so far
The timeline is short, so here it is.
- September 22, 2026, Apsara Conference, Hangzhou. Alibaba said its next-generation Qwen4 was in training and put four names on screen: Qwen4-Max, Qwen4-Plus, Qwen4-Flash and Qwen4-27B. No dates, prices or specs came with them.
- October 29, 2026, Apsara Conference, Amsterdam. A slide read "Qwen4 Series Coming Soon" over the same four names. An attendee posted that Alibaba Cloud CTO Feifei Li said "next week", which would mean November 2 to 8.
- Since then. Alibaba has not repeated that date on its own channels. Several trackers say plainly that no model card, license or downloadable weights exist.
That "next week" line is the thing to be careful with. It comes second-hand from someone in the room. It may well be right. But "coming soon" is the phrase companies use when they do not want to put a date in writing, and I would not plan a project around a sentence overheard at a conference.
The four Qwen4 variants, one by one
| Variant | Role (as described) | Open weights? | Specs published? |
|---|---|---|---|
| Qwen4-Max | Flagship, hardest tasks | Unconfirmed | No |
| Qwen4-Plus | Balanced, multimodal tier | Unlikely (Plus has never been open) | No |
| Qwen4-Flash | Fast, low-cost, high volume | Unconfirmed | No |
| Qwen4-27B | Local deployment tier | Described as open, not yet released | Size only (27B in the name) |
Qwen4-Max
The flagship. Code changes spotted in Alibaba's Qwen Code repository reference a Qwen4-Max identifier, and a separate change reportedly adds support for it in kernel code for Alibaba's own accelerators. That tells you engineers are working on it. It does not tell you the size, the price or whether you will ever download it.
Qwen4-Plus
Plus is the odd one. Alibaba apparently skipped a standalone Plus in the 3.8 generation, so seeing the name return is mildly interesting. Historically it has been an API-only tier, and nothing suggests that changes.
Qwen4-Flash
The efficient tier. Its architecture is the one people can partly inspect today, through a preview model I will get to below.
Qwen4-27B
This is the one most developers are searching for, because it is the one you could run yourself. Alibaba has framed it as the open-weights, local-deployment model. Its size is in the name, but whether it is a dense model or mixture-of-experts has not been stated.
Why "0 open weights confirmed" is the real story
Here is the distinction that gets lost. Saying a model is meant to be open is not the same as releasing weights, and neither is the same as publishing a license. Developers care about the third part most, because a license decides whether you can build a product on the thing.
An open-weights label without a license text is a headline, not a plan.
Right now we have the first part at best. No downloadable file exists for any Qwen4 model, and Alibaba has not said under what terms one would arrive. At the time of writing, a public prediction market on whether Alibaba publishes Qwen4 weights by November 25 was sitting at roughly 42 percent. Markets are noisy, and I would not lean on that number. But it shows even people with money on the line are not treating this as settled.
The clue that is already public: Qwen3.8-Flash-Next
In late August 2026, Alibaba released Qwen3.8-Flash-Next with open weights and described it as an experimental preview of the architecture that will underpin Qwen4. Reported figures put it at roughly 125 billion parameters in total with about 6 billion active per token, a 51 billion parameter n-gram embedding component, and a native context window of 262,144 tokens, extendable to about one million.
If you want to know how Qwen4 might behave, this is the closest thing you can actually run. The catch is that Alibaba's own framing, as relayed by trackers, is that Flash-Next specs belong to the 3.8 generation and should not be copied onto any Qwen4 tier. Treat it as a hint about direction, not a spec sheet.
Alibaba's track record on openness is mixed
This is where my skepticism comes from. Qwen is widely described as a mostly open-weights family, and Hugging Face downloads of Qwen models have passed 700 million, according to one report. But the most advanced models have often stayed closed.
- The Qwen3.7 models, Max in May and Plus in June 2026, were proprietary.
- Qwen3.8-Max reached the cloud on August 3, and its weights followed on August 12 under a different name, with some features such as image input and a non-thinking mode left out.
- Reporting around the Qwen team has also pointed to a leadership exodus, which is one reason some outlets are questioning how firm the open-source pledge really is.
So there is precedent both ways. Alibaba has opened a flagship before, and it has kept one closed before. That is exactly why "unconfirmed" is the honest word.
My take from the tech desk
I have been on the technology beat for a year, which is not long, but it is long enough to see the pattern. A lab names a model, the community fills the silence with specs, and by the time the real release arrives half the "facts" in circulation turn out to have been guesses. I would rather be the boring reporter who says "we do not know yet" than the one who repeats a leaked parameter count.
What stands out to me about Qwen4 is the gap between volume and substance. There is an enormous amount of writing about it and a very small amount of verified material. Four names. One conference slide, shown twice. A handful of code identifiers. That is the whole confirmed pile.
One more thing I find telling: the biggest number in the Hangzhou coverage was not about Qwen4 at all. It was a comment from Alibaba CEO Eddie Wu about a later model in the 5 to 10 trillion parameter range. Alibaba is already talking past Qwen4. If you are a developer, that is a reminder to build around interfaces, not around a single model name.
What developers and readers should do this week
- Do not rebuild anything yet. Keep production on the shipped generation until Alibaba publishes a model card and a license.
- Test the architecture early. Qwen3.8-Flash-Next is downloadable, so you can see how the new design handles your workload.
- Keep your serving layer swappable. If you call models through one compatible interface, switching to Qwen4 later is a config change, not a rewrite.
- Read the license before celebrating. "Open weights" can still carry commercial limits.
- Trust official channels. Alibaba's Qwen team page and its Hugging Face organization are the places a real release will show up first.
Conclusion
Qwen4 is real in the sense that matters least: it has a name, a slide and some code identifiers. It is not real in the sense that matters most, because nobody outside Alibaba can download, license or benchmark it. Four variants named, zero open weights confirmed. That is the accurate summary today.
The 27B is the one to watch, and the license text is the part to read first. Until then, plan for the models that exist.
Frequently asked questions
What is Qwen4?
Qwen4 is the next generation of Alibaba's Qwen language models. Alibaba showed four names at its Apsara conference: Qwen4-Max, Qwen4-Plus, Qwen4-Flash and Qwen4-27B. As of November 3, 2026, none has been released.
Will Qwen4 have open weights?
Not confirmed. Qwen4-27B has been described as the open-weights tier, but Alibaba has not published weights, a license or a model card for any Qwen4 model.
When is the Qwen4 release date?
Alibaba has not announced one. An attendee at Apsara Amsterdam on October 29 reported that Alibaba Cloud CTO Feifei Li said "next week", which points to November 2 to 8, 2026. That is unofficial.
Which Qwen4 model is most likely to be open?
Qwen4-27B. Max, Plus and Flash are expected to be API models, and Plus has never shipped as open weights.
Can I try the Qwen4 architecture today?
Partly. Qwen3.8-Flash-Next, released in late August 2026 as open weights, is described as an experimental preview of the Qwen4 architecture.
Is there a Qwen4 coding model?
Not officially. Identifiers such as qwen4-coder-plus appear in Qwen Code's repository, but Alibaba has announced no coding variant.