US Edition
Your source for latest news

Alibaba Launches Qwen4-Max, Betting Lower Prices Can Challenge US AI Leaders

Alibaba's newest flagship model cuts API prices 25 percent and claims a 62 percent drop in per-task cost versus its predecessor, while the company says it deliberately stripped offensive cyberattack and bioweapons-related capability from its training data.

PT
By PressTemps Technology DeskPublished Today, 00:45 ET · 5 min read
Alibaba Launches Qwen4-Max, Betting Lower Prices Can Challenge US AI Leaders
Alibaba Group's Taobao City campus in Hangzhou, China, photographed in 2014. Alibaba's Qwen AI team, based primarily in Hangzhou and Beijing, released the Qwen4-Max model on November 16, 2026. Credit: Danielinblue / Wikimedia Commons (CC BY-SA 4.0).
What to know
Alibaba released Qwen4-Max, a 2.4-trillion-parameter model activating 71 billion parameters per token, with a 1-million-token context window
API prices fell 25 percent to $1.50/$4.50 per million input/output tokens; Alibaba claims a 62 percent drop in average cost per task on its internal Work360 benchmark
The model stays closed (no public weights), unlike predecessor Qwen3.8-Max, though Alibaba says smaller open-weight Qwen4 models are coming soon
Alibaba says it filtered offensive cyberattack data from training and added CBRN safety post-training, part of wider industry debate over open-weight model risk

Alibaba Group released Qwen4-Max on Monday, the newest flagship model in its Qwen artificial-intelligence family, positioning the system as a cheaper and more capable rival to the closed models sold by OpenAI and Anthropic. The launch, detailed in a post on Alibaba's Qwen blog, is the company's most direct challenge yet to American AI labs on cost and coding performance, even as it keeps the model itself closed rather than releasing its underlying weights.

Qwen4-Max is a mixture-of-experts system with 2.4 trillion total parameters, the same overall size as its predecessor, Qwen3.8-Max, but it activates only 71 billion parameters per token, about 25 percent fewer than before. Alibaba says that leaner activation pattern is what allows the new model to reason more efficiently without growing the compute bill. The model accepts text, image and video input, carries a one-million-token context window, and is documented in detail on Alibaba Cloud's Model Studio reference page, which lists a maximum combined output of 262,144 tokens across its reasoning and final-answer stages.

Lower prices, lower per-task cost

Alongside the model, Alibaba cut API prices. According to Alibaba Cloud's published rate card, Qwen4-Max costs $1.50 per million input tokens and $4.50 per million output tokens in its Singapore-hosted international pricing, down from $2 and $6 for Qwen3.8-Max, a 25 percent reduction on both ends. Cached input reads cost as little as 7.5 cents per million tokens. Alibaba also published results from an internal benchmark called Work360, which it says measures performance on realistic office and engineering tasks; it reported the average cost of completing a Work360 task fell from $4.43 to $1.67, a drop of roughly 62 percent. Those figures come from Alibaba's own testing and have not been independently reproduced.

Independent benchmarking also offers an early, more modest read. Artificial Analysis, a firm that evaluates AI models on a standardized battery of reasoning, coding and math tests, scored Qwen4-Max at 43 on its Intelligence Index, placing it above the median of the roughly 80 models it tracks but behind several top closed systems from Anthropic, OpenAI and Google.

A closed model from an open-weight champion

The release is notable in part because of what Alibaba chose not to do. Qwen4-Max itself remains closed, available only through Alibaba's cloud platform, in contrast to Qwen3.8-Max, whose weights Alibaba released openly in August. The company says it still plans to open-source smaller models in the Qwen4 line "soon," continuing a strategy that has made Qwen one of the most widely used open-weight model families among developers worldwide, competing domestically against DeepSeek and Moonshot's Kimi models as well as against OpenAI, Anthropic and Google abroad.

That strategy underscores how central large language models have become to Alibaba's broader cloud-computing business, which has leaned on Qwen's popularity to court enterprise customers in Singapore, Germany, Japan, Hong Kong and the United States, the five international regions where Qwen4-Max is now deployed alongside mainland China. The new model also arrives against the backdrop of a three-year, roughly 380-billion-yuan ($53 billion) commitment to AI and cloud infrastructure that Alibaba Group announced in February 2025, which it described at the time as exceeding the company's combined AI and cloud spending over the previous decade. That buildout, still underway, is meant to supply the computing capacity behind successive Qwen releases and to support Alibaba Cloud's push for enterprise customers outside China, where it competes with Amazon Web Services, Microsoft Azure and Google Cloud as well as with homegrown rivals.

Cybersecurity and bioweapons guardrails

Alibaba used the launch to address a growing concern among AI-safety researchers: that increasingly capable models, especially ones released with open weights, could make it easier for malicious actors to carry out cyberattacks or design biological, chemical, radiological or nuclear weapons. The company said it filtered offensive cyberattack material out of Qwen4-Max's training data while retaining data for defensive tasks such as vulnerability detection and remediation, and that it designed dedicated safety data for CBRN risks during post-training.

"No capabilities posing large-scale cybersecurity risks have been observed," Alibaba said in its Qwen4-Max release notes.

The company added that internal reviewers found the model's proposed solutions to CBRN-related prompts were "not genuinely innovative" and could not complete such a task end-to-end, identifying no risk of the model translating relevant knowledge into a significant uplift in a user's capabilities. Those are Alibaba's own internal assessments rather than findings from an outside evaluator, and the broader question of whether open-weight models carry greater CBRN risk than the closed systems favored by American labs such as OpenAI and Anthropic remains an active and unresolved debate among researchers, who have flagged the possibility that future, more capable models could meaningfully lower the technical bar for building dangerous weapons.

Who is affected, and what comes next

The immediate audience for Qwen4-Max is developers and enterprises already using Alibaba Cloud's Model Studio platform, plus coding tools and agent workflows that Alibaba says can call the model through standard OpenAI- and Anthropic-style APIs. Alibaba highlighted a handful of internal case studies alongside the launch, including a claim that a 48-hour autonomous research run identified a candidate coolant molecule for next-generation semiconductor chips, and that an automated legal-review workflow organized nearly 200 due-diligence documents into an 86-page report. None of those results have been verified by outside researchers.

For the broader AI industry, the launch adds to a steady drumbeat of large model releases from Chinese developers this year, reinforcing a competitive dynamic in which Alibaba, DeepSeek and Moonshot have each released frontier-scale systems within months of one another, often undercutting US rivals on price even when trailing them on independent benchmark scores. Alibaba said smaller, open-weight versions of the Qwen4 family are coming "soon," which would extend the pattern set by Qwen3.8-Max of eventually opening up models that launch behind a paywall. No firm date has been given for those releases, nor for Qwen4-Max-Preview's scheduled retirement on November 25, a limited-access version Alibaba had been testing with developers since November 11.

Coverage of the launch in trade press, including an analysis published by Unite.AI, noted that Alibaba has restricted the steep Token Plan subscription discount to interactive coding and agent-tool use through products such as Qoder and Qwen Code, excluding bulk automation scripts and non-interactive batch workloads from the cheaper rate. That distinction matters for enterprise customers evaluating the true cost of adopting Qwen4-Max at scale, since list prices on the pay-as-you-go API remain the figures businesses running automated pipelines would actually pay. Alibaba has not said when, or whether, it will extend Token Plan pricing to those use cases.

More on this story

All Technology