Brief IA
›

Alibaba Launches Qwen4-Max with 2.4 Trillion Parameters and Paid API

Code & Dev·Tom Levy·

Alibaba Launches Qwen4-Max with 2.4 Trillion Parameters and Paid API

Alibaba Launches Qwen4-Max with 2.4 Trillion Parameters and Paid API
Key Takeaways
1Qwen4-Max accessible via API in QwenCloud, at $1.50/M tokens for input and $4.50/M for output
2Mixture-of-experts architecture: 2.4 trillion parameters, 71 billion activated per token, context announced at 1 million tokens
3Alibaba reports up to 62% lower costs and a reduction in tokens on Work360 and Terminal-Bench
Why it matters : commercial access, public pricing, and the lack of open weights necessitate evaluating Qwen4-Max on real workloads, beyond Alibaba's internal tests.

Alibaba launches Qwen4-Max, a giant mixture-of-experts model announced with 2.4 trillion parameters. The company boasts significantly reduced costs in its internal tests while commercializing access via API at public rates and not publishing the weights.

Access via QwenCloud, public rates, and non open-weight model

The official version of Qwen4-Max is available within the Qwen ecosystem, particularly on QwenCloud, with an API accessible based on usage pricing. The international price announced is $1.50 for one million input tokens and $4.50 for one million output tokens, excluding special situations related to caching and billing. Unlike other Qwen models, Alibaba has not made the weights of Qwen4-Max public. The company plans to offer smaller Qwen4 models with open weights, although no timeline has been provided. At this stage, developers need to evaluate the service, its performance, security, and data handling. The preview version was initially reserved for the Token Plan before broader access was opened. Announced on November 11, the Qwen4-Max-Preview had a limited access window until November 25.

Architecture with 2.4 trillion parameters, 71 billion activated per token

Qwen4-Max is announced with 2.4 trillion parameters, similar to Qwen3.8-Max, but relies on a mixture-of-experts architecture that activates only about 71 billion parameters for each processed token. The context window is reported to be one million tokens. The model accepts text, images, and videos as input and produces text responses. A large window allows for more elements to be provided, without guaranteeing uniform attention or the absence of errors.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Cost reductions claimed by Alibaba and lower token usage

Alibaba reports tasks being up to 62% less expensive in its own tests. On Work360, which corresponds to its internal evaluation tool, the average cost per task was $4.43 with Qwen3.8-Max, then decreased to $1.67 using Qwen4-Max. The company explains these results by a reduction in token consumption. According to its measurements, Qwen4-Max uses about 24 to 35% fewer tokens on Work360, and 48 to 49% on Terminal-Bench 4.0 depending on the settings. However, these cost reductions are not guaranteed for all types of usage. The final cost varies based on the nature of the request, interaction with tools, and applied rates, and this data must be compared to independent workloads.

Demonstrated cases: game engines, Blender, due diligence, and agents

Alibaba indicates that Qwen4-Max is capable of controlling programming tools, creating playable prototypes in Unity, Unreal Engine, or Godot, and also presents examples of 3D modeling done with Blender. For a legal due diligence case presented by Qwen, the model reportedly structured 198 documents and generated an 86-page report. Sequences of analysis and experimentation are also detailed on partner platforms, requiring multiple calls to tools. The company pursues an agent-oriented goal, through Qwen-AgentWorld and the training of AI agents. These demonstrations do not guarantee that all users will achieve the same results. Qwen4-Max was officially announced on November 16, 2026, following a preview available on November 11. Alibaba has not demonstrated that it has created the best AI on the market, but it advocates for an AI capable of performing more tasks without skyrocketing costs for businesses. The model will need to prove itself on real tasks, beyond scoreboards.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.