Alibaba's Qwen4-Max lowers costs and strengthens reasoning

Alibaba opens Qwen4-Max to the public after a preview reserved for subscribers. The model promises lower costs, better reasoning, and enhanced safeguards, in an increasingly fierce global competition.

Alibaba made Qwen4-Max publicly accessible on November 16, 2026, after an open preview on November 11, with availability announced at 3:01 a.m. Paris time. This AI model targets professional and scientific uses while promising a 62% reduction in average cost per task. In a market where every performance point is monetized, the figure attracts as much attention as the caution it demands.

A giant model, but designed to activate only part of its parameters

Qwen4-Max is based on a Mixture-of-Experts (MoE) architecture, which activates only a fraction of the model per request. Alibaba claims to have 2.4 trillion parameters in total, with 71 billion activated per token, 25% less than Qwen3.8-Max-Base. This choice aims to reduce computational load without sacrificing processing capacity, a compromise that has become central for large-scale models.

The model also includes native visual understanding and a context window that can reach one million tokens. In practice, this allows it to absorb long documents, analyze images, and handle more complex requests than simple text exchanges. Alibaba highlights use cases in 3D modeling, programming, industrial design, and scientific research.

Cheaper to use, more aggressive on the market

On its internal Work360 benchmark, covering 721 specialized domains, Alibaba claims that Qwen4-Max outperforms Qwen3.8-Max. In advanced reasoning mode, the model would consume 35% fewer tokens, according to the company. However, these results had not yet been confirmed by independent evaluations at the time of launch [to be verified].

Competitive pressure also comes through pricing. Qwen4-Max is priced at $1.50 per million input tokens and $4.50 per million output tokens, compared to $2 and $6 for Qwen3.8-Max. According to Alibaba's tests, the average cost per task would thus drop from $4.43 to $1.67. This reduction changes the game for companies industrializing AI, as costs quickly add up when scaling the model.

Enhanced safeguards, but still to be tested

Alibaba claims to have added security controls during training to limit vulnerabilities related to code injections, memory security flaws, and access control issues. The group also says it excluded certain data associated with cyberoffensive activities while retaining data related to vulnerability detection and correction.

Specific mechanisms have also been integrated to reduce risks related to chemical, biological, radiological, and nuclear uses. However, their effectiveness will need to be validated by independent tests; otherwise, the security promise will remain mostly a promise.

An offensive that goes beyond Alibaba alone

With Qwen4-Max, the Chinese group positions itself against OpenAI, Anthropic, Google, and Chinese players like DeepSeek. The strategy is clear: combine performance, low costs, and more visible security to attract companies. In this context, the choice will not be based solely on raw power but on reliability, data protection, and ease of integration.

For buyers, this multiplication of models broadens the offer but also complicates comparisons. Alibaba's announcements reinforce a clear trend: the AI battle is now fought as much on price as on response quality. And sometimes, the token saved is almost as valuable as the token generated.

Key points

  • Qwen4-Max has been publicly accessible since November 16, 2026.
  • Alibaba announces a total of 2.4 trillion parameters.
  • 71 billion parameters are activated per token.
  • The average cost per task decreases by 62%, according to Alibaba.
  • Work360 covers 721 specialized domains.
  • The model targets finance, health, law, and scientific research.

By the numbers

  • 62% — decrease in average cost per task announced by Alibaba.
  • 2.4 trillion — total parameters of Qwen4-Max.
  • 71 billion — parameters activated per token.
  • $1.50 — price per million input tokens.
  • $4.50 — price per million output tokens.

To read