Haitong International has released a research report indicating that MINIMAX-W (00100) has built a strong foundation in computing infrastructure, general model architecture, and multimodal products. The firm believes the company is poised for a period where capability upgrades and revenue growth will reinforce each other.

According to the brokerage's research, the company's next-generation model is moving toward a parameter scale of approximately 3 trillion. It will focus on optimizing the intelligence output per unit of computing power through techniques such as Mixture of Experts (MoE), sparse attention, KV Cache compression, and reinforcement learning. The native multimodal approach will also be tested with the launch of the next Hailuo product.

Observations and tracking show that the domestic large model sector has not yet developed stable user lock-in or a clear leader. Inference efficiency, post-training quality, product experience, and real-world scenario penetration are expected to be the key decisive factors in the next phase. Although the industry still faces risks from technological iteration and intensifying competition, the company's long-term growth potential is worth noting. Haitong maintains an "Outperform" rating for the company.

The report points out that MINIMAX-W places a high priority on the intelligence output per unit of computing power in its model development. It reduces token costs through architectural design, self-built computing capacity, and collaborative optimization of inference software. The M3 model faced margin pressure during its initial discounted launch, but after continuous optimization, its gross margin has significantly improved. This suggests that the company's pursuit of cost-effectiveness is not simply about low prices but involves maintaining model capability and inference speed while using restrained pricing to lower the entry barrier for customers. The company then relies on call volume and ongoing cost reduction to achieve reasonable returns.

Based on observations and tracking, as model capabilities become more aligned, the capability-to-price ratio will directly influence enterprise customers' decisions on API migration and application expansion. If MINIMAX-W can maintain a faster rate of cost reduction than the market price decline, its cost-effectiveness could become a key lever for expanding market share.

According to the brokerage's research, MINIMAX-W's ARR (Annual Recurring Revenue) increased from approximately $100 million at the end of 2025 to over $150 million by February 2026. By April 2026, it had again achieved significant growth compared to February, with a full-year target of $1 billion. The rapid revenue growth indicates that the model's capabilities have already formed a commercial feedback loop through consumer-grade products and enterprise APIs. The report suggests that future focus should be on tracking the actual performance and pricing of the approximately 3T model, user retention and paid conversion for the new Hailuo product, improvements in inference gross margin, and progress in overseas computing and compliance.