GTHT Securities has issued a research report maintaining an "Overweight" rating on MINIMAX-WP (00100) with a target price of HKD 1012. The firm made minor adjustments to the company's revenue forecasts for 2025-2027, now projecting USD 0.7 billion, USD 2.2 billion, and USD 5.5 billion respectively. Considering MINIMAX-WP's status as a rare global multimodal large language model player listed in Hong Kong, coupled with the strong catalyst from the recent launch of its flagship M2.5 model, the report applies a 186x price-to-sales ratio for 2026. Key viewpoints from GTHT Securities are outlined below.
The company's technological capabilities have achieved breakthroughs in both performance and efficiency, placing its core scenario performance in the global top tier. The MiniMax M2.5 model, supported by optimized Mixture of Experts architecture and the native Agent RL framework Forge, delivers state-of-the-art industry performance in productivity scenarios such as programming, tool usage, search, and office tasks. Its core metrics are comparable to leading overseas models like Claude Opus 4.5. Furthermore, the model's inference speed reaches 100 tokens per second, doubling the industry mainstream level. Its processing speed on the SWE-Bench Verified task has improved by 37% compared to the M2.1 version, demonstrating dual breakthroughs in performance enhancement and efficiency optimization.
Exceptional cost control has established a competitive barrier based on cost-effectiveness, making commercial deployment economically viable. Through token consumption optimization, upgraded parallel tool calling, and a tiered pricing design, M2.5 has built an industry-leading cost advantage, breaking the economic barriers to large-scale operation of complex agents. The pricing strategy employs a dual-version approach: the 100 TPS fast version costs only USD 1 for one hour of continuous operation, while the 50 TPS version is priced as low as USD 0.3. Calculated by output price, this is merely 1/10th to 1/20th the cost of overseas models like ClaudeOpus, Gemini 3 Pro, and GPT-5, positioning it as a global benchmark for high-value large language models.
Product iteration and ecosystem development are driving growth simultaneously, securing a key position in the emerging era of universal agents. MiniMax demonstrates an industry-leading model iteration pace, having completed three generational updates (M2, M2.1, M2.5) within the past 108 days. This rate of progress significantly outpaces overseas giants like Anthropic, OpenAI, and Google, primarily attributable to the deep synergy between large-scale reinforcement learning and engineering capabilities. Concurrently, the M2.5 model has been fully deployed on the MiniMax Agent platform, which distills core Office Skills capabilities and allows users to create reusable industry experts. The platform has already generated over 10,000 custom experts, covering high-frequency scenarios such as office work, finance, and programming.
The report highlights several risk factors, including slower-than-expected commercial progress, industry competition and technological iteration risks, compliance and legal risks, and cross-market valuation risks.