Source : Alibaba Cloud
HANGZHOU, November 16, 2026 — Alibaba has launched Qwen4-Max, the new flagship model in its Qwen family of artificial intelligence systems, saying it can handle text, images and video, tackle complex reasoning and coding, and do so at a markedly lower cost per task than the previous generation.
According to Alibaba Cloud Model Studio documentation updated on November 16, Qwen4-Max is a multimodal model — meaning it can take more than one kind of input — combining text generation with image and video understanding and deep reasoning. The company said it is designed for coding, office work, frontend design and “agentic” tasks, in which the model carries out a sequence of steps, including web searches and tool calls, rather than only answering a single question. It also supports structured outputs and is compatible with the OpenAI and Anthropic application programming interfaces, so developers can plug it into existing tools. The documentation lists a context window of 1,000,000 tokens, the units a model uses to process text, and says the model is available in regions including Beijing, Singapore, Frankfurt, Virginia, Tokyo and Hong Kong.
A limited preview, Qwen4-Max-Preview, was opened on November 11, the same documentation shows. The November 16 release is the full model. In a company blog post described by IT Home, Alibaba’s Qwen team said Qwen4-Max uses a sparse mixture-of-experts design, in which only part of the network runs for each request, with a total of 2.4 trillion parameters. The model is not open-weight, meaning users cannot download and customise it themselves, The Straits Times reported, citing Bloomberg; Alibaba said smaller models in the Qwen4 series would be made available as open-source later.
Alibaba’s own performance claims should be read as the company’s figures. It said Qwen4-Max scored near the top of coding benchmarks and outperformed Anthropic’s Claude Fable 5.1 and Moonshot’s Kimi K3.1 Pro, according to The Straits Times. On a professional-work test of its own, called Work360, the Qwen blog said the new model scored 52.69, compared with 47.74 for Qwen3.8-Max, IT Home reported. The same account said the average cost of a Work360 task fell to 11.16 yuan from 26.58 yuan, a drop of 58 per cent, after both lower token use and lower API prices. Token consumption at comparable performance fell by about 24 per cent and 35 per cent at two higher reasoning settings, and by about 48 per cent and 49 per cent on Terminal-Bench 4.0, the blog said.
The company has also emphasised safety. The South China Morning Post reported that Alibaba said it filtered the model’s training data to remove offensive cyber-attack capabilities while keeping abilities useful for cyber defence, such as spotting software vulnerabilities, and that it introduced data sets and extra training to reduce chemical, biological, radiological and nuclear risks.
Developers can call the model through Alibaba’s API and Token Plan, IT Home reported, and users can try it in Qwen Studio, QwenWork and Qoder. A preview had already been limited to Token Plan customers.
What to watch next is whether independent tests back Alibaba’s coding and cost claims, when the promised smaller open-weight Qwen4 models actually appear, and how the tighter security limits affect what the model will and will not do in the hands of developers.
Source : Alibaba Cloud Model Studio, “Model releases and updates” and “Qwen4-Max”, November 16, 2026; Alibaba’s Qwen blog, as reported by IT Home; The Straits Times (Bloomberg); and the South China Morning Post.

