Alibaba Cloud CTO: Qwen4 Lands Next Week in Max, Plus, Flash, 27B

Model launches usually come with a carefully staged announcement. Qwen4 may have been let out of the bag by a single line on stage. On October 29, at Alibaba Cloud's first European Apsara Conference, held in Amsterdam's old stock exchange, the Beurs van Berlage, CTO and international president Feifei Li told the room: "Qwen4 next week." Someone in the audience typed it straight onto X.

Later, Qwen researcher Xudong Guo took the stage in front of a slide reading "Qwen4 Series Coming Soon," with four names lined up underneath: Qwen4-Max, Qwen4-Plus, Qwen4-Flash and Qwen4-27B. Read literally, "next week" means somewhere between November 2 and 8.

The official line is still "in training"

Notably, Qwen4 was absent from Alibaba's formal announcements that day. The press release covered the company's first cloud region in the Netherlands, its fourth in Europe after Frankfurt, London and Paris, plus Qwen3.8-Flash going live on Model Studio in the EU. The @Alibaba_Qwen account has not confirmed the date.

The four model names themselves are not new. At the Hangzhou Apsara in September, Alibaba said Qwen4 was in training and sketched a roadmap toward Qwen 4.5 and Qwen 5 at 5 to 10 trillion parameters, but gave no dates, prices or weights. August's Qwen3.8-Flash, a multimodal MoE with 125 billion parameters plus 51 billion N-gram embeddings, was pitched as a preview of the Qwen4 architecture.

So the only genuinely new thing in Amsterdam is a timeline. The stage line said "next week"; the slide said "coming soon." The gap between those is the gap between a commitment and an intention. Still, the Qwen team has a habit of shipping with little more than a post on X and weights already sitting on Hugging Face, so "next week" is not a long shot.

The repo hints at a bigger family

Around the same time, developers spotted a PR merged the day before in Qwen's official repository that references more variants: on top of the four on the slide, qwen4-coder-plus and qwen4-omni-120b-think, plus a qwen4.5-max. All of the IDs map to a 262,144-token context, and the omni ID is already in use as a session model, which makes them look less like placeholders.

If that list holds, Qwen4 is a whole-family refresh rather than a single upgrade: Max and Plus on the API to take on closed flagships, Flash for price and speed, Coder for agentic programming, Omni for multimodal reasoning. It is Alibaba's playbook from the past year: fill every price tier so that a developer at any budget finds a Qwen.

27B is the one developers actually care about

Community chatter quickly narrowed to one concrete question: will 27B run on a single 3090 or 5090?

That makes sense. Max, Plus and Flash will most likely live in the cloud, while 27B is the one that looks like the open, local model. Rule of thumb: a dense 27B at 4-bit is a little over 15GB of weights. That fits on a 24GB 3090 but leaves little headroom for long-context KV cache, while a 32GB 5090 is far more comfortable. The real answer waits on weights and architecture details. If 27B follows Qwen3.8-Flash into MoE or extra embeddings, the memory math changes.

Seen that way, 27B is Qwen4's crowd-pleaser. Local deployment has lived in the 20B–35B band for the past year, and whoever delivers the best quality on one consumer GPU becomes the default in Ollama, LM Studio and countless local agent projects. Qwen already has a strong reputation there; Qwen4-27B has to defend it against pressure from Gemma, Mistral and DeepSeek's smaller models.

Alibaba needs a good launch

Dropping the line at its European debut shows Alibaba knows how much Qwen matters abroad. For many European developers, their first impression of Alibaba Cloud is not a region but the Qwen weights near the top of Hugging Face's download charts. Warm up a new region with "see you next week," then follow through with a family launch: it is a natural one-two punch.

What to watch over the next week: whether weights ship at launch, the license and quantized builds for 27B, Max's benchmarks and API pricing, and whether that omni-120b-think shows up too. If "next week" holds, early November in open models is going to be Qwen's.