Alibaba launches Qwen4-Max with a one-million-token context

Alibaba launches Qwen4-Max with a one-million-token context

Qwen4-Max supports a one-million-token context and processes text, images, and video. Alibaba introduced it on November 16 as its new flagship model and said it achieved programming results above those of Anthropic's Claude Fable 5.1 and Moonshot's Kimi K3.1 Pro.

A multimodal model with a one-million-token context

Alibaba describes Qwen4-Max as the next-generation flagship multimodal model in the Qwen family. In addition to generating text, it can understand images and videos; its output is text. The company targets it at programming, office productivity, interface design, and complex agent-based tasks.

The context limit is 1,000,000 tokens. In non-thinking mode, the maximum input is 991,808 tokens and the maximum output is 131,072. In thinking mode, input can reach 983,616 tokens; the maximum combined total for reasoning and the final response is 262,144 tokens.

ParameterLimit
Context1,000,000 tokens
Maximum input, without reasoning991,808 tokens
Maximum input, with reasoning983,616 tokens
Maximum output, without reasoning131,072 tokens
Combined reasoning and final response262,144 tokens

Alibaba Cloud's specifications also list web search, tool and function calls, parallel calls, structured JSON responses, and compatibility with the OpenAI and Anthropic interfaces, in addition to DashScope. It supports streaming responses and encrypted reasoning summaries. The company thus proposes using it in applications that combine multimodal understanding with task execution.

The comparison with Anthropic and Moonshot

In its announcement, Alibaba said Qwen4-Max ranked near the top on programming benchmarks and outperformed Claude Fable 5.1 and Kimi K3.1 Pro. The company presented it as a competitive response to Anthropic and OpenAI. The published information does not provide numerical results from those tests or a complete ranking, so the comparison is limited to Alibaba's claim.

The launch comes amid a succession of announcements from Chinese companies seeking to challenge the United States' lead in artificial intelligence. The Straits Times names Alibaba, Moonshot, and DeepSeek among them; the article also places Qwen4-Max's arrival after the recent launches of Tencent's Hunyuan 4 and Moonshot's Kimi K3.1 Pro.

The Qwen family was already, according to The Straits Times, the world's most downloaded among AI models with open weights. However, Qwen4-Max will not have open weights: users will not be able to download and customize its parameters as they can with models released that way. Alibaba said it will soon release smaller versions of the series, but did not specify when or what features they will have.

Access from six regions and usage-based pricing

Alibaba Cloud Model Studio offers the model under the identifier qwen4-max. The documentation lists deployments in Beijing, Singapore, Frankfurt, Virginia, Tokyo, and Hong Kong. Beijing is listed as limited to mainland China; Singapore as an international service; and the other regions as global.

RegionListed coverageInput per million tokensOutput per million tokens
BeijingMainland China10 yuan30 yuan
SingaporeInternational10.938 yuan32.813 yuan
FrankfurtGlobal10 yuan30 yuan
VirginiaGlobal10 yuan30 yuan
TokyoGlobal10 yuan30 yuan
Hong KongGlobal10 yuan30 yuan

The rates apply to requests of up to one million tokens and are the same for thinking and non-thinking modes. The Beijing page lists a free quota of one million tokens, valid for 90 days from the activation of Model Studio, the model's release, or the application's approval, whichever date is later. The documentation does not show this quota in the tables for other regions.

Usage limits are also presented differently by region. For Virginia, Frankfurt, Tokyo, and Hong Kong, Alibaba specifies up to 30,000 requests per minute and five million tokens per minute, counting both inputs and outputs; it notes that per-second limits may apply. For Beijing and Singapore, it publishes tokens-per-minute tiers in Bailian: five million in the first tier, ten million in the second, and twenty million in the third, depending on the platform's monthly spending.

The scheme assigns the first tier to monthly spending of up to 100,000 yuan, the second to between 100,000 and one million, and the third to more than one million. The specifications do not allow model deployment or fine-tuning, although they do indicate support for streaming, explicit and implicit caching, and structured formats.

Security and Alibaba's ecosystem

Alibaba incorporated enhanced security measures into Qwen4-Max following a series of incidents that, according to The Straits Times, fueled concerns about the cybersecurity risks of models with open weights. The outlet adds that Tencent and Moonshot adopted similar measures. It does not describe the incidents or detail the specific controls Alibaba added.

The decision comes as the company seeks to build an AI system broader than a collection of models: its plan includes models, artificial intelligence chips, and greater computing capacity. Qwen's momentum is tied to a strategy led by CEO Eddie Wu, according to The Straits Times.

In September, Wu said that T-Head, Alibaba's chip manufacturing unit, would launch China's most powerful AI chips for the domestic market. T-Head had recently filed for an initial public offering in Hong Kong. The Hangzhou-based company has made artificial general intelligence its primary goal, and Wu set a target of 20 GW of global data center capacity by 2032.

The company's next announced step is the arrival of smaller versions of the Qwen family.

Sources: The Straits Times, Alibaba Cloud