Alibaba has brought to the generative AI market a new flagship model Qwen4-Max, which is becoming a key element of the Qwen family and sets a new scale for enterprise language systems. The development demonstrates that Chinese tech giants continue to increase the computational power of their models, despite global competition and restrictions on the supply of advanced hardware.
The key feature of Qwen4-Max is its Mixture of Experts (MoE) architecture. The total number of model parameters reaches 2.4 trillion, yet only 71 billion parameters are activated to process each individual token. This approach allows for a significant reduction in computational load during inference while maintaining access to the model's vast aggregate capacity.
Why the MoE architecture matters
The MoE mechanism routes each token to a limited set of "expert" subnetworks, providing a balance between performance and cost efficiency. This very principle underlies most of the latest generation of advanced models and allows systems to be scaled to trillions of parameters without a proportional increase in computational cost.
For Alibaba, Qwen4-Max is not only a technological but also a strategic asset: the model strengthens the company's position in the cloud segment and enterprise AI services, where the deployment of efficient MoE architectures is becoming a competitive advantage.
Cryptalist analysis: The emergence of a model with 2.4 trillion parameters while activating only 71 billion confirms the steady trend toward "sparse" architectures as the main direction of generative AI development. For the market, this is a signal of intensifying competition between Chinese and Western providers, which in the long term will put pressure on inference costs and stimulate demand for computing infrastructure and related digital assets.