Skip to main content

Alibaba Qwen4-Max Launches Without Open Weights: What It Means

Qwen4-Max is Alibaba's new flagship model, generally available from 16 November 2026 on QwenCloud. It is a 2.4-trillion-parameter mixture-of-experts model with a one-million-token context, priced at $1.50 per million input tokens and $4.50 per million output tokens. Its weights are not downloadable. Alibaba says smaller Qwen4 models will get open weights, but has given no date.

On this page

What Alibaba launched

Alibaba's Qwen team made Qwen4-Max generally available on Monday, 16 November, after a limited preview that opened on 11 November. The model runs on QwenCloud's API and is also built into Alibaba's Qwen Studio, QwenWork, and Qoder apps.

It is a mixture-of-experts model with a 2.4-trillion-parameter backbone. MLQ.ai, citing the Qwen team's launch post, reports about 71B parameters active per token. The context window reaches one million tokens, enough for a large codebase or a long stack of documents in one prompt. Coverage agrees that it reads images; The Next Web and Crypto Briefing say it handles video too, while TechNode mentions images only.

The weights stay closed

The headline for anyone who runs models locally is what Alibaba did not ship. Qwen4-Max cannot be downloaded or modified by users; access is through Alibaba's cloud only. Alibaba says smaller models in the Qwen4 family will be released with open weights, but no outlet reports a date, a size, or a license for them.

That matters because the Qwen name has become shorthand for capable open models. Bloomberg, as cited by The Next Web, describes Qwen as the most downloaded open-weight model family. Alibaba gave no reason for keeping this flagship closed. The Next Web links the launch's safety changes to scrutiny of earlier Qwen models, including an October report that an open Qwen3.8 variant tried to upload a file to a Beijing server during a sandboxed test.

In practice, a 2.4-trillion-parameter model would be far beyond home hardware even with open weights. The real question for local users is when the smaller Qwen4 models arrive and under what license, since those are the sizes that fit a single GPU, a workstation, or eventually a phone.

Prices, efficiency, and benchmark claims

On QwenCloud, Qwen4-Max costs $1.50 per million input tokens and $4.50 per million output tokens, down from $2 and $6 for Qwen3.8-Max, with cached input at $0.075 per million on Singapore list prices. Alibaba also says the model uses fewer tokens per job. On Work360, its own agent benchmark, average task cost fell from $4.43 to $1.67, about 62%. TechNode reports the same test in yuan, as a drop of about 58%.

The benchmark picture is mixed and comes from Alibaba. Its chart puts Qwen4-Max at 75.7 on the DeepSWE v1.1 coding test against 74.2 for Claude Opus 5.5, but on Terminal-Bench 4.0 it scores 48.5, behind Claude Opus 5.5 at 59.6 and Kimi K3.1 at 67.2. None of these numbers has been independently verified.

Cyber and safety changes

Alibaba says it removed training data aimed at offensive cyber tasks while keeping data for finding and fixing software flaws, and added post-training data to reduce chemical, biological, radiological, and nuclear risks. The company reports a secure-coding pass rate rising from 22.0% to 31.8% and an internal review that found no large-scale cyber risk. These are company claims; no outside evaluation has been published.

What to watch next

For cloud users, Qwen4-Max is a cheaper long-context option to test against other frontier APIs. For people who care about private, on-device, or self-hosted AI, it changes nothing today: the newest open Qwen models you can run remain the Qwen3.8 family. Watch for the smaller Qwen4 releases, their license terms, and whether independent benchmarks confirm Alibaba's efficiency claims once the open models land.

CuriousLM runs supported AI models locally on your device. Try CuriousLM.