Haitong International has released a research report stating that MINIMAX-W (00100) has built deep expertise in computing infrastructure, general model architecture, and multimodal products. The firm believes this positions the company for a convergence of capability upgrades and revenue growth. According to the brokerage's research, the company's next-generation model is moving toward approximately 3 trillion parameters, with optimizations in MoE, sparse attention, KV cache compression, and reinforcement learning to enhance the intelligence output per unit of computing power. The native multimodal roadmap will also be tested with the launch of Hailuo's new generation products.

Observations and tracking show that the domestic large model market has not yet formed stable user lock-in or a clear leader. Inference efficiency, post-training quality, product experience, and penetration into real-world scenarios will become the key differentiating factors in the next phase. Despite the risks of technological iteration and intensifying competition, the company's long-term growth potential is worth attention, leading Haitong to maintain its 'Outperform' rating for the company.

The report indicates that MINIMAX places a strong emphasis on the intelligence output per unit of computing power in its model development. It reduces token costs through architectural design, self-built computing infrastructure, and collaborative optimization of inference software. The M3 model faced pressure on profitability during its initial launch with promotional pricing, but gross margins have improved significantly after continuous optimization. This suggests that its pursuit of cost-effectiveness is not simply about low prices, but rather about keeping pricing disciplined to lower the barrier to entry for customers while maintaining model capability and inference speed, ultimately achieving reasonable returns through call volume and ongoing cost reduction.

According to observations, as model capabilities gradually converge, the capability-to-price ratio will directly influence the migration and application expansion of enterprise customers. If MiniMax can maintain a faster rate of cost reduction than the market price decline, its cost-effectiveness could become a key lever for expanding its market share.

Based on the firm's research, MiniMax's ARR has risen from approximately $100 million at the end of 2025 to over $150 million by February 2026, and experienced significant growth again from February to April 2026, with a full-year target of $1 billion. This rapid revenue growth indicates that the model's capabilities are already forming a commercial closed loop through consumer-grade products and enterprise APIs. The report suggests that subsequent tracking should focus on the actual performance and pricing of the approximately 3T model, user retention and paid conversion for Hailuo's new products, improvements in inference gross margins, and progress in overseas computing and compliance.