AI

Alibaba launches Qwen4-Max to take on OpenAI and Anthropic. What it claims, and what it means for Australians

By Nino Ray Yeh · November 16, 2026 · 7:52 pm AEDT · 6 min read
The Alibaba Group headquarters in Hangzhou, with white lattice-patterned facades, a matching lattice canopy and glass sky bridges between the buildings

Alibaba has released Qwen4-Max, the new flagship of its Qwen family of AI models, and it is pitching the model directly at the top-tier systems from OpenAI and Anthropic. In its announcement, dated Monday 16 November in Hangzhou, the company calls Qwen4-Max “the most capable model in the Qwen family to date” and says it is built for coding, AI agents and long-running professional and scientific work.

The launch is the latest in a rapid run of Chinese model releases, and it comes with an unusually prominent section on safety. Here is what Alibaba has actually announced, which of its claims have independent backing (and which don’t yet), and what it means for Australian developers and businesses weighing up AI tools.

What Alibaba announced

According to Alibaba, Qwen4-Max is a Mixture-of-Experts (MoE) model with 2.4 trillion total parameters, of which 71 billion are active for each token it processes. In plain terms, only a fraction of the model does the work on any given request, which is meant to keep running costs down. Alibaba says it has native vision capabilities (it can work with images as well as text) and a one-million-token context window, meaning it can take in very large documents or codebases at once.

The company says Qwen4-Max is “available now on QwenCloud via the Token Plan and API”, as well as through its QwenWork and QwenStudio platforms. It lists one price in the announcement: cached input at $0.075 per million tokens (the currency isn’t specified). Reuters reported the launch, as did Bloomberg, via The Straits Times and Lianhe Zaobao.

Importantly, the flagship is not an open-weight model, so you can’t download it and run it on your own hardware. The Straits Times report notes “Alibaba did not make the model open weight”, while Alibaba says “Open-weight, smaller-scale models from the Qwen4 series will soon be released to developers worldwide.” No release date has been given for those smaller models.

The performance claims, and what backs them

Nearly every performance number in the launch comes from Alibaba itself. The company says Qwen4-Max ranked 6th on Code Arena, a public coding leaderboard, and posted strong results on several agent, workplace and visual-reasoning benchmarks. On Work360, which Alibaba describes as a proprietary benchmark covering 721 specialised fields, it says the new model beats its predecessor, Qwen3.8-Max, while using 35% fewer tokens in its high-effort reasoning mode and cutting the average cost per task by around 60%.

Reports differ on how Alibaba framed the comparison with US rivals. Bloomberg’s report says the model outperformed Anthropic’s Claude Fable 5.1 and Moonshot’s Kimi K3.1 Pro on coding benchmarks. Mint reported a coding score of 75.7 against 75.2 for OpenAI’s GPT-6.1-Sol and 74.2 for Anthropic’s Opus 5.5, and noted that the figures “are based on Alibaba’s reported benchmark results and have not been independently verified.” That caveat applies to all of the benchmark claims here: a 6th-place leaderboard ranking and one or two points on a vendor-selected test are not proof the model is better for your work.

Alibaba also highlights demonstrations rather than independent tests: it says Qwen4-Max cut manual programming time for a manufacturing workpiece from roughly 475 minutes to 24 minutes in a showcase, and completed a continuous 48-hour agent run that found an experimentally validated chip coolant molecule. Those are company accounts, and the announcement doesn’t include enough detail for outsiders to reproduce them.

A bigger emphasis on safety

Alibaba devotes a section of the announcement to safety, saying software engineering training “included code security checks alongside functionality tests” to reduce vulnerabilities such as injection flaws, memory-safety bugs and broken access control in the code the model writes. It says it filtered cybersecurity training data “by task intent”, keeping defensive material for finding and fixing vulnerabilities and removing offensive attack content, and added safety datasets and training aimed at chemical, biological, radiological and nuclear (CBRN) risks.

Reuters reported Alibaba’s statement that internal testing found no evidence of significant high-risk capabilities. That is a self-assessment: Alibaba hasn’t published an independent evaluation, and the announcement doesn’t name an outside tester. Reuters also notes that Chinese AI developers have increased safety disclosures in recent months “amid growing global concern surrounding “rogue AI” and loss-of-control risks”.

Why it matters

Qwen matters beyond China because of its open models. Bloomberg describes the Qwen series as “the most downloaded open-weight AI model family in the world”, and many developers build on Qwen models because they can be downloaded, run locally and customised. If the promised smaller Qwen4 models arrive with similar efficiency gains, they could become a popular free alternative for businesses that want to keep AI on their own servers. We covered Alibaba’s wider ambitions, including its own AI chips, in our report on its V900 chip.

It is also another sign that the gap between Chinese and US frontier models is narrow, at least on the companies’ own numbers. That competition has come with friction: the US has previously accused DeepSeek, Alibaba and other Chinese AI firms of large-scale model copying.

What it means in Australia

For Australian government agencies, there is a specific rule for one Chinese AI provider, but not this one. Home Affairs’ PSPF Direction 001-2025 requires Australian Government entities to “prevent the use or installation of DeepSeek products, applications and web services” and remove existing instances from government systems and devices. That direction names DeepSeek only; it doesn’t mention Alibaba or Qwen.

For businesses, Alibaba’s announcement doesn’t say where QwenCloud processes or stores data for overseas customers, or whether Qwen4-Max is offered through any Australian-hosted service. That matters if you’re thinking of sending customer or staff information to it. A few practical points:

  • The OAIC’s guidance on commercially available AI products says privacy obligations apply to personal information put into an AI system, and that “the OAIC recommends that organisations do not enter personal information, and particularly sensitive information, into publicly available generative AI tools”.
  • Check where a provider processes and stores data, how long it keeps prompts and outputs, and whether they’re used for training, before connecting any AI service to business systems. That applies to US and Chinese providers alike.
  • Treat vendor benchmarks as a starting point. If you’re comparing Qwen4-Max with Claude, GPT or Gemini models, test them on your own (non-sensitive) tasks and measure cost per task, not just price per token.
  • If you’re interested in running models locally, keep an eye out for the smaller open-weight Qwen4 releases, and check their licence terms when they arrive.

Our read

Qwen4-Max looks like a serious release, and the efficiency claims (fewer active parameters and lower cost per task) are the part most likely to matter if they hold up. But almost everything we know about its performance comes from Alibaba, and the most interesting part for many developers, the smaller open-weight Qwen4 models, hasn’t shipped yet. The new safety section is welcome, but a company marking its own homework isn’t the same as independent testing. We’ll be watching for third-party evaluations, the open-weight release, and clearer answers on where customer data is handled before we’d suggest Australian businesses treat it as a like-for-like swap for the models they already use.

Featured image: Alibaba Group headquarters in Hangzhou, China, 2012. Photo: Thomas LOMBARD (building designed by HASSELL), via Wikimedia Commons, licensed CC BY-SA 3.0; cropped to 3:2, resized and converted to PNG. Illustrative only; the photo is not connected with the Qwen4-Max announcement.

Keep exploring

Read next

More stories worth your time from The Tech Boom.

Explore all tech stories
Share this story

Topics

Share with