ThreadAI: capabilities, regulation, labour

AI & Tech · Beijing

Alibaba unveils Qwen4-Max, a closed flagship AI model pitched against OpenAI and Anthropic

The closed-weight model costs 67% less per task than its predecessor, Qwen3.8-Max, and provisionally ranks fifth overall on the Arena.ai leaderboard, behind four US firms.

AI text based on reports by pb.pl, Notícias ao Minuto, EL MUNDO and other publishers. Published . How it is made

For membersThe first 30 seconds for everyone. Synthetic voice (AI).

In brief

Beijing · China

Launch and benchmark claims

Alibaba presented Qwen4-Max on Monday, 16 November, as its new flagship AI model, positioned against systems from OpenAI and Anthropic. According to Alibaba, the model scored near the top of coding benchmarks and outperformed Anthropic's Claude Fable 5.1 and Moonshot's Kimi K3.1 Pro. The company says it also improves on its predecessor, Qwen3.8-Max, which was unveiled in August, while using fewer tokens. Alibaba describes gains in coding, reasoning and complex professional work, and lists uses that range from 3D modelling and game development to finance, law, healthcare and scientific research. The release is one of a run of Chinese model launches, alongside Tencent's Hunyuan 4 and Moonshot's Kimi K3.1 Pro, from companies that are challenging the lead of US AI firms.

Size, rankings and cost

Qwen4-Max is a closed model with 2.4 trillion parameters in its main architecture, a figure close to the 2.8 trillion attributed to Kimi K3. On the Arena.ai evaluation platform, it was provisionally placed first among Chinese models and fifth overall, behind systems from Anthropic, Google, OpenAI and Meta. Artificial Analysis puts the cost at $3.27 per task on its intelligence index, 67% below the $10.02 for Qwen3.8-Max. The reduction results from both lower prices and lower token consumption, tokens being the units AI systems use to process and generate responses.

Parameters in main architecture of flagship modelstrillion parameters
Qwen4-Max
2.4 trillion parameters
Kimi K3
2.8 trillion parameters

Data from the article text

Cost per task on Artificial Analysis intelligence indexUSD per task
Qwen3.8-Max
10.02 USD per task
Qwen4-Max
3.27 USD per task

Data from the article text

Cybersecurity safeguards

Alibaba says it filtered the training data to remove capabilities that could support cyberattacks, while keeping data used for defensive work such as finding and fixing software vulnerabilities. Tests run during training targeted code injection, memory management errors and flawed access controls. The company also says it strengthened protections against the use of AI for cybercrime and for creating dangerous material. Additional post-training measures aim to mitigate chemical, biological, radiological and nuclear risks, according to the company. Alibaba's internal review concluded that the model's cyberdefence capabilities improved substantially and that protection against offensive use was also significantly reinforced. These safety claims come from Alibaba's own statement.

Access and infrastructure plans

Qwen4-Max is not released as an open-weights model, so users cannot download it and modify it themselves. Alibaba has said that smaller versions of the Qwen family will be made available. Bloomberg reports that the Qwen family is the most downloaded open-weights model family in the world, a result it attributes to a large-scale push led by group chief executive Eddie Wu. Alibaba is building a full AI ecosystem that runs from models to its own AI chips. The company aims to reach artificial general intelligence and plans data centres with a combined capacity of 20 GW by 2032.

This article is part of a thread

AI: capabilities, regulation, labour

What changed since this entry:

  1. 16/11Alibaba unveils Qwen4-Max, 67% cheaper per task than predecessor, fifth on Arena.ai
  2. 16/11Anchor adds the 15 November US-led AI supply-chain declaration with eight partners

Ask about this text

Sources and context

Pollar is free and carries no ads in the text. Its members keep it going.

Support Pollar