Qwen4 arrives: Alibaba opens preview of flagship Qwen4-Max

What's happening

Chinese tech giant Alibaba has officially launched its new Qwen4 model generation. First out of the gate is a preview of the flagship Qwen4-Max. The Qwen team announced it on November 11, 2026, via its official X account (@Alibaba_Qwen), according to Japanese tech outlet ITmedia AI+. You can try the model through Alibaba Cloud's "Token Plan" subscription service. Early access is also available in Alibaba's own apps: the office app QwenWork (called "Qwen Office" in Chinese reports), the coding tool Qoder, the Qwen app and QwenStudio.

One thing up front: Qwen4-Max is a cloud model. The weights aren't available for download, and the model runs only on Alibaba's servers. Many Qwen fans will still be paying close attention, because the Qwen4 family is also supposed to include a version you can run on your own machine.

What Qwen4-Max is supposed to do

According to ITmedia, Alibaba hasn't published any technical details yet. There's no word on model size and no official benchmarks. The company says, however, that the model has improved at understanding knowledge and the real world, at coding, and at working through tasks on its own. The announcement reads:

Built for real-world tasks – from complex coding and knowledge work to frontend development and professional software workflows.

Chinese business outlets ChainCatcher and HuoXing Finance (via KuCoin) point to three focus areas: complex coding, frontend design (the visible interface of websites and apps) and agentic capabilities. An AI agent is a model that doesn't just answer but operates tools itself – launching programs, editing files or searching the web, for example – to complete multi-step tasks. Alibaba also highlights office work and handling specialist software. In other words, the model should be able to deal with spreadsheets, documents and professional applications too.

Months in the making: the new architecture

Qwen4 doesn't come out of nowhere. According to ChainCatcher, Alibaba announced in September that Qwen4 was being trained on a new architecture. A preview of that came earlier with Qwen3.8-Flash-Next, an experimental open-weights release. Tech site The Next Web (TNW) took a closer look:

• Qwen3.8-Flash-Next has 125 billion parameters in total but activates only 6 billion per generated token (word fragment). That's thanks to a Mixture-of-Experts design: for each input, only a few specialized subnetworks do the work, not the whole model.
• For comparison, the older Qwen3.7-Plus has 397 billion parameters, 17 billion of them active. So the new model needs only about a third of the compute per token.
• A new form of "sparse attention" selects small blocks of tokens rather than individual tokens. The text is processed in selected chunks instead of in full, which saves compute on long inputs.
• A "gated residual" mechanism acts like a valve, controlling which information gets passed between the network's layers.
• An unusual addition is a separate 51-billion-parameter lookup module. Instead of adding more experts, the team bolts on a kind of dictionary that's queried via two- or three-character sequences. According to the team, this costs less compute and is easier to offload to chips with limited memory.

TNW reads that last point politically: if you're designing for accelerators with tight memory, it's usually because you can't get the best chips. That's exactly where Chinese labs find themselves under US export controls. TNW says the team is less focused on raw performance than on lower costs: AI agents with very long contexts are becoming a standard workload, and the inference bills are turning into a problem.

The Qwen4 family: Flash, Plus and a 27B version

According to the roadmap already shown, the lineup will include Flash and Plus variants alongside Max, plus a 27B version for local use. For tinkerers, that last one is the most interesting. A model this size fits on well-equipped gaming GPUs with quantization (storing the model's numbers at lower precision so it needs less memory). Alibaba hasn't confirmed when the variants will ship or whether the 27B version will again be freely downloadable.

First comparison: a 27B model on par with Sonnet 5.5

An early impression comes from a post on the blogging platform note.com about the private "GameFeel Bench" benchmark, which tests AI models on game development. It covered early development builds that the tester calls Qwen4. The author flags a caveat: the name and size of "Qwen-4-27B" are the tester's own labels. The Qoder team also hasn't said what's behind the mysterious test model "Qoduck", which the tester links to it.

Overall scores after a correction (the 27B model was initially reported at 47.0 points; a calculation error in the audio scoring brought it down to 44.6):

"XHigh" and "Max" are different levels of how long the model is allowed to think. Head to head against Sonnet 5.5 XHigh, the 27B model leads in six of seven categories: implementing new features (64 vs. 63), finding bugs (63 vs. 62), game feel and balance (51 vs. 48), ideas (43 vs. 39), audio (28 vs. 20) and reviewing other models' proposed changes (14.88 vs. 14.26). Only in graphics does Sonnet come out ahead (53 vs. 48).

Against Sonnet 5.5 at the Max level, the picture changes. Qwen wins just three categories and falls well behind in reviewing others' proposals (14.88 vs. 23.33). The author cautions against reading too much into the numbers. The 1.8-point gap to Sonnet XHigh isn't statistically sound, since there are no error bars or repeated runs. And the 49.9 points belong to the Flash version and shouldn't be credited to the 27B model.

Licensing questions for Europe

TNW highlights an issue that could matter for future open Qwen4 models. Qwen3.8-Flash-Next is on Hugging Face under a "qwen-community" license, and according to TNW, Alibaba plans to charge its largest commercial users. Under the EU AI Act, lighter documentation requirements apply only to models released under a free and open-source license. Per Recital 103, paid or otherwise monetized components are meant to be excluded. Anyone building on Qwen – TNW cites Thomson Reuters as an example – inherits the license in whatever form it ends up taking.

The big picture

With Qwen4-Max, Alibaba is back in the frontier-model race, aiming squarely at the area where the most money is flowing right now: coding and agentic work. How good the model really is remains to be seen. There are no official benchmarks, no technical specs and no independent tests of the Max model itself. TNW notes that Alibaba already called Qwen3.8 the "second-best model in the world" without evidence, and that the numbers usually come from its own test setups.

For many readers, the announced 27B version is the more exciting part. If it really ships openly and comes close to Claude Sonnet 5.5 in practice, that would be a big step for local AI. For now, though, that hope rests on a single private benchmark of an unconfirmed preview build. The timeline, the license and the real-world performance are all still open questions.