高盛在最新研报中认为MiniMax是中国人工智能模型公司中布局最为出色的企业之一，在文本与编程领域具备显著的全球潜在市场（TAM）增长 空间。 据追风交易台，2月23日，高盛最新研报首次覆盖中国全模态AI公司MiniMax，给予"中性"评级，基于DCF模型设定的12个月目标价为1018港元， 对应企业估值约389亿美元（约合2820亿元人民币）。 高盛指出，MiniMax作为全球顶尖的AI模型公司之一，正处于高速增长阶段。其最大的亮点在于高度全球化的收入结构（70%的收入来自海外） 以及全面的多模态产品线（涵盖文本、视频、音频、音乐和图像）。高盛指出，MiniMax在文本/代码、多模态以及AI智能体/数字劳动力领域的广 阔全球潜在市场（TAM）中占据了极佳的竞争位置。 对投资者的核心影响： 多模态优势与全球化商业版图 极致的成本效率与底层架构创新 凭借MoE架构和线性注意力机制，以极低成本逼近全球顶尖水平。 高盛估计，MiniMax仅投入约4.2亿美元（2022-2025年预测）的训练成本，就跻身全球顶级AI公司之列。这得益于其早期采用的混合专家（MoE） 架构和线性注意力机制，大幅降低了计算资源消耗和推理成 ...

在行业分布上，中证人工智能主题指数(930713)覆盖AI全产业链，从底层算力、中层技术到上层应用， 完整覆盖人工智能三大核心环节。 指数前十大重仓股，按AI细分赛道分类来看，光模块/通信占比21.14%（中际旭创、新易盛），AI芯片/ 处理器占比14.51%（寒武纪、澜起科技、豪威集团），AI服务器/算力占比4.70%（中科曙光），算法/ 大模型/NLP占比4.36%（科大讯飞），计算机视觉占比4.11%（海康威视），AI应用/IP平台占比5.22% （金山办公、芯原股份）。 规模与流动性方面优势显著，截止2025年12月31日，天弘中证人工智能主题指数基金产品规模达33.94 亿元，A类：7.80亿，C类：26.14亿（数据来源产品2025年四季报）。 基金业绩方面，2025年历史收益亮眼：天弘中证人工智能主题指数基金A类：64.98%，C类：64.66% （数据来源产品2025年四季报）。 国内方面，2026年春节前后国产大模型迎来集体发布潮，智谱发布GLM-5大模型，参数达744B，编程 能力对标Claude Opus 4.5，且完成与多款国产算力平台的适配；MiniMax推出M2.5模型，在编程、工 ...

Core Viewpoint - Cathay Securities maintains an "overweight" rating for MINIMAX-WP (00100) with a target price of 1012 HKD, slightly adjusting the company's revenue forecast for 2025-2027 to 0.7/2.2/5.5 billion USD, considering its unique position as a global multimodal large model in the Hong Kong stock market and the strong catalyst from the recent M2.5 flagship model release, assigning a 186x PS for 2026 [1] Group 1: Technical Capabilities - MiniMax M2.5 leverages MoE architecture optimization and the native Agent RL framework Forge, achieving industry-leading performance in programming, tool invocation, search, and office productivity scenarios, with core metrics comparable to top overseas models like Claude Opus 4.5 [1] - The model's inference speed reaches 100 TPS, which is double the industry average, and SWE-Bench Verified task processing speed has improved by 37% compared to M2.1, demonstrating a dual breakthrough in performance enhancement and efficiency optimization [1] Group 2: Cost Control and Commercial Viability - M2.5 establishes a leading cost advantage through token consumption optimization, parallel tool invocation upgrades, and tiered pricing design, overcoming economic barriers for large-scale Agent operations [2] - The pricing strategy includes a dual-version approach, with a 100 TPS fast version costing only 1 USD for continuous operation for one hour, and a 50 TPS version priced as low as 0.3 USD, making it 1/10 to 1/20 the cost of overseas models like Claude Opus, Gemini 3 Pro, and GPT-5, positioning it as a global benchmark for high cost-performance large models [2] Group 3: Product Iteration and Ecosystem Development - MiniMax demonstrates industry-leading model iteration speed, completing updates for M2, M2.1, and M2.5 within 108 days, significantly outpacing overseas giants like Anthropic, OpenAI, and Google, thanks to deep synergy between large-scale reinforcement learning and engineering capabilities [3] - The M2.5 model is fully integrated into the MiniMax Agent platform, refining core Office Skills capabilities and enabling users to create reusable industry experts, with over 10,000 custom experts now available covering high-frequency scenarios in office, finance, and programming [3]

Summary of AI Industry Dynamics During the 2026 Spring Festival Domestic Model and Product Releases 1. **Zhiyuan AI** (February 11): Released GLM-5 with a HumanEval code pass rate of 96.2%, ranking first in global open source and fourth overall; focuses on programming and intelligent capabilities [1] 2. **ByteDance** (February 14): Launched Doubao Model 2.0 (Pro/Lite/Mini/Code), achieving top rankings in math/programming benchmarks and reducing inference costs by an order of magnitude; simultaneously launched Seedance 2.0, capable of generating movie-quality multi-shot audio and video in 60 seconds, becoming a key visual creation tool for the Spring Festival Gala [1] 3. **Alibaba** (February 16): Open-sourced Qwen3.5-Plus with a total parameter count of 397 billion and only 17 billion activated, achieving a 60% reduction in memory usage and a 19-fold increase in inference efficiency; API priced at 0.8 yuan per million tokens, which is 1/18 of Gemini 3 Pro [1] Additional Domestic Releases 4. **DeepSeek** (February 17): Upgraded context window from 128K to 1M tokens, capable of processing ultra-long texts equivalent to the "Three-Body Problem" trilogy [2] 5. **MiniMax** (February 18): Released M2.5 model with native agent design, achieving a 37% speed increase for complex tasks [2] 6. **Tencent** (February 19): AI creation reached 1 billion instances over 16 days during the Spring Festival [2] Overseas Model and Product Releases 7. **Google** (February 19): Released Gemini 3.1 Pro, significantly enhancing inference capabilities and setting a new technical benchmark [3] 8. **OpenAI** (February 18): Launched a proprietary model based on Cerebras chips, achieving a training efficiency increase of 5-10 times and a cost reduction of 70% [3] 9. **Anthropic** (February 18): Released Claude Sonnet 4.6, priced at 1/5 of flagship models, with performance approaching Opus, offering excellent cost-performance for enterprise applications [3] 10. **xAI** (February 20): Updated Grok 4.2, with simultaneous upgrades in multimodal and inference capabilities [3] Financing and Capital Dynamics 11. **Moon's Dark Side (Kimi)** (February 20): Completed a $700 million financing round led by Alibaba, Tencent, and Wuyuan, with a valuation exceeding $10 billion; cumulative financing in January and February surpassed $1.2 billion [4] 12. **Zhiyuan AI/MiniMax** (February 20): Market capitalization on the first trading day after the Hong Kong holiday exceeded 300 billion HKD each, totaling over 580 billion HKD [4] 13. **Anthropic** (February 21): Completed a $30 billion Series G round, with a valuation reaching $380 billion, led by GIC and Coatue [4] 14. **OpenAI** (February 21): Nearing completion of the first phase of over $100 billion in financing, with a valuation expected to exceed $850 billion [4] 15. **Runway** (February 21): AI video company completed a $315 million Series E round, with a valuation of $5.3 billion, focusing on world model development [4]

Core Insights - Chinese technology companies are launching new AI models, with Alibaba's Qwen3.5-Plus being a significant highlight, featuring 397 billion parameters and a cost-effective API pricing of 0.8 yuan per million tokens, which is 1/18 of Gemini3 Pro's price [1][2] - The trend in AI model development is shifting from sheer size to efficiency and intelligence, as demonstrated by Qwen3.5's ability to perform well with a smaller model while maintaining high performance [2][3] Group 1: Model Launches and Features - Alibaba launched the Qwen3.5-Plus model on New Year's Eve, which has 397 billion total parameters and 170 billion activated parameters, achieving a 60% reduction in deployment memory usage [1] - Other companies like Zhiyun, iFlytek, and MiniMax have also introduced new models, including GLM-5, Xinghuo X2, and M2.5, showcasing advancements in decision-making capabilities [1][2] Group 2: Performance and Efficiency - Qwen3.5 has shown exceptional performance in various benchmark tests, achieving results comparable to Gemini3 Pro and excelling in visual understanding assessments [1][2] - The new models are designed to be more practical and efficient, with a focus on multi-modal capabilities and reduced resource requirements, leading to increased deployment efficiency [2][3] Group 3: Open Source and Global Impact - The number of open-source models related to Qwen has exceeded 400, with over 200,000 derivative models and downloads surpassing 1 billion, indicating a strong global presence [3] - China is emerging as a leading provider of open-source large models, with significant contributions from companies like Qwen and DeepSeek, which rank highly on AI model evaluation platforms [3]

Core Insights - The AI sector in Hong Kong has shown strong performance, particularly in areas such as large models, storage, and computing power, driven by a wave of new product launches ahead of the Lunar New Year [1][3] Group 1: Product Launches and Innovations - Alibaba is set to release its new generation Qwen 3.5 large model on New Year's Eve, featuring architectural innovations [3] - ByteDance has launched the Doubao large model 2.0, which has undergone significant optimization for complex tasks [3] - Zhipu has introduced its flagship model GLM-5, increasing its parameter scale from 355 billion to 744 billion, with an average performance improvement of over 20% [3] - MiniMax has released the M2.5 model, designed for agent scenarios, achieving a 37% improvement in task completion speed compared to its predecessor [3] - Other products like DeepSeek-V4 and Kimi-K3 are expected to be released within February, indicating a significant product launch wave in the domestic large model sector [3] Group 2: Market Dynamics and Future Outlook - More domestic AI companies are choosing to list in Hong Kong, injecting new vitality into the sector, exemplified by Haizhi Technology Group's stock surging over 320% shortly after its listing [3] - Zhongyuan Securities predicts that the pace of AI application deployment in China will exceed market expectations by 2026, with domestic large models set to rival overseas counterparts in terms of performance, power consumption, and pricing [3] - Bank of America Securities notes that the acceleration of large model iterations in China is enhancing data center demand, which will drive adoption among enterprises and developers, subsequently increasing demand for inference data centers [3][4] Group 3: Structural Opportunities in AI Computing - Guosheng Securities highlights that breakthroughs from companies like ByteDance and Alibaba are shifting the focus of AI applications from technology development to large-scale deployment, creating a strong demand for AI computing resources [4]

Market Overview - US CPI growth slowed to 2.4% year-on-year in January, below expectations, leading to reduced inflation concerns and increased expectations for interest rate cuts, with traders now estimating a 50% chance of three rate cuts this year [10][28] - The S&P 500 index experienced a weekly decline of 1.39%, while the Nasdaq fell 2.1%, marking significant market volatility [17] - In Asia, the Shanghai Composite Index fell below 4100 points, with semiconductor stocks performing well, while the Hang Seng Index dropped by 1% [2] Key Developments in China - The People's Daily published an important article by General Secretary Xi Jinping emphasizing the need to boost domestic demand and implement measures to stimulate consumption and investment [23] - China's social financing in January reached 7.22 trillion yuan, with new RMB loans amounting to 4.71 trillion yuan, and M2 money supply growing by 9.0% year-on-year [24] - The housing market showed mixed signals, with a narrowing month-on-month decline in prices across first, second, and third-tier cities, but an expanding year-on-year decline [4][24] Regulatory Actions - The Financial Regulatory Bureau and the Market Regulatory Bureau held discussions with six travel platform companies regarding compliance issues in their lending practices [5][25] - New antitrust compliance guidelines were issued, highlighting eight new types of monopolistic risks, including "choose one" and "lowest price across the network" practices [25] Company News - Meituan forecasted a significant loss of over 23 billion yuan for 2025, continuing losses into the first quarter [9][25] - DeepSeek is expected to release a new model, V4, around the Spring Festival, which is currently undergoing testing [6][26] - ByteDance launched Doubao 2.0, significantly reducing inference costs and competing directly with GPT-5 and Gemini 3 [7][27] - MiniMax introduced the M2.5 model, which operates at a cost of $1 per hour, significantly cheaper than GPT-5, while maintaining competitive performance [8][26] International Developments - The US is considering partially lifting tariffs on aluminum and steel to alleviate inflationary pressures [13][29] - The US military is preparing to deploy a second aircraft carrier to the Middle East amid rising tensions with Iran [14][41] - Trump confirmed plans to visit Venezuela, with US energy officials noting that Venezuelan oil revenues have exceeded $1 billion [15][30]

Core Viewpoint - MiniMax has launched its latest M2.5 model series, achieving a significant breakthrough in both performance and cost, aiming to address the economic feasibility of complex agent applications while claiming to have reached or refreshed the industry SOTA (state-of-the-art) levels in programming, tool invocation, and office scenarios [3][4]. Cost Efficiency - The M2.5 model demonstrates a substantial price advantage, costing only 1/10 to 1/20 of mainstream models like Claude Opus, Gemini 3 Pro, and GPT-5 when outputting 50 tokens per second [3][4]. - In a high-speed environment of 100 tokens per second, the cost for continuous operation for one hour is just $1, and it can drop to $0.3 at 50 tokens per second, allowing a budget of $10,000 to support four agents working continuously for a year [3][4]. Performance Metrics - M2.5 has shown strong performance in core programming tests, winning first place in the Multi-SWE-Bench multi-language task, with overall performance comparable to the Claude Opus series [4]. - The model has improved task completion speed by 37% compared to the previous generation M2.1, with an end-to-end runtime reduced to 22.8 minutes, matching Claude Opus 4.6 [4]. Internal Validation - Internally, MiniMax has validated the M2.5 model's capabilities, with 30% of overall tasks autonomously completed by M2.5, covering core functions such as R&D, product, and sales [4]. - In programming scenarios, M2.5-generated code accounts for 80% of newly submitted code, indicating high penetration and usability in real production environments [4]. Task Efficiency - M2.5 aims to eliminate cost constraints for running complex agents by optimizing inference speed and token efficiency, achieving a processing speed of 100 TPS (transactions per second), approximately double that of current mainstream models [7]. - The model has reduced the total token consumption per task to an average of 3.52 million tokens in SWE-Bench Verified evaluations, down from 3.72 million in M2.1, allowing for nearly unlimited agent construction and operation economically [9]. Programming Capability - M2.5 emphasizes not only code generation but also system design capabilities, evolving a native specification behavior that allows it to decompose functions, structures, and UI designs from an architect's perspective before coding [11]. - The model has been trained in over 10 programming languages, including GO, C++, Rust, and Python, across tens of thousands of real environments [12]. Testing and Validation - M2.5 has been tested on programming scaffolds like Droid and OpenCode, achieving pass rates of 79.7% and 76.1%, respectively, outperforming previous models and Claude Opus 4.6 [14]. Advanced Task Handling - In search and tool invocation, M2.5 exhibits higher decision maturity, seeking more streamlined solutions rather than merely achieving correctness, saving approximately 20% in rounds consumed compared to previous generations [16]. - For office scenarios, M2.5 integrates industry-specific knowledge through collaboration with professionals in finance and law, achieving an average win rate of 59.0% in comparisons with mainstream models, capable of producing industry-standard reports, presentations, and complex financial models [18]. Technical Foundation - The performance enhancement of M2.5 is driven by large-scale reinforcement learning (RL) through a native Agent RL framework named Forge, which decouples the underlying training engine from the agent, supporting integration with any scaffold [23]. - The engineering team has optimized asynchronous scheduling and tree-structured sample merging strategies, achieving approximately 40 times training acceleration, validating a near-linear improvement in model capabilities with increased computational power and task numbers [23]. Deployment - M2.5 is fully deployed in MiniMax Agent, API, and Coding Plan, with model weights to be open-sourced on HuggingFace, supporting local deployment [25].

Investment Rating - The report maintains an "Overweight" rating for the computer industry, expecting the industry index to outperform the market index by over 10% in the next six months [30]. Core Insights - The report highlights the recent surge in AI-related activities, particularly the "Spring Festival Treat Plan" by Alibaba's Qianwen APP, which led to a significant increase in AI-driven orders, indicating a growing demand for AI computing power [3][27]. - The report emphasizes the potential investment opportunities in the AI computing power supply chain, driven by the increasing adoption of AI applications by major tech companies [3][27]. Summary by Sections 1. Industry Performance Review - The SW computer sector experienced a cumulative decline of 0.49% over the past two weeks, outperforming the CSI 300 index by 0.23 percentage points, ranking 20th among 31 sectors [11]. - In February, the sector rose by 1.50%, surpassing the CSI 300 index by 1.22 percentage points, and has increased by 8.55% year-to-date, outperforming the CSI 300 index by 6.61 percentage points [11][16]. 2. Valuation Situation - As of February 12, 2026, the SW computer sector's PE TTM (excluding negative values) stands at 59.40 times, placing it in the 96.39th percentile over the past five years and the 91.04th percentile over the past ten years [21][22]. 3. Industry News - Key developments include the launch of the Seedance 2.0 video generation model by Doubao, which enhances video content creation capabilities [22]. - Alibaba's Qianwen APP achieved over 1.2 billion AI-driven orders within six days of launching its promotional campaign, highlighting the rapid growth in AI application usage [22]. - Major tech companies, including Amazon, Google, Microsoft, and Meta, are projected to increase their capital expenditures to $660 billion in 2026, a 60% increase from 2025 [22]. - MiniMax introduced its M2.5 model, significantly reducing operational costs while maintaining competitive performance [22]. - Google is integrating AI shopping features into its search engine, allowing direct purchases through AI-driven interactions [22]. 4. Company Announcements - Notable company announcements include: - Guiding Compass reported a 40.39% increase in revenue for 2025, reaching 2.146 billion yuan [25]. - Zhongke Shuguang plans to raise up to 8 billion yuan through convertible bonds for AI-related projects [25]. - Unisplendour announced a plan to raise up to 5.57 billion yuan for acquisitions and R&D [26]. 5. Weekly Perspective - The report suggests continued monitoring of Alibaba and other AI giants' strategies in the AI application space, as their growth is expected to drive demand for computing power [27]. 6. Recommended Stocks - The report lists several companies to watch, including: - GuoDianYunTong (002152.SZ) for its stable growth in fintech and AI sectors [28]. - Digital China (000034.SZ) as a key player in the domestic computing demand [28]. - Inspur Information (000977.SZ) for its leadership in AI server markets [28].

Core Insights - MiniMax has launched its latest M2.5 series model, significantly reducing inference costs while maintaining industry-leading performance, aiming to address the economic feasibility of complex agent applications [1] - The M2.5 model demonstrates a substantial price advantage, costing only 1/10 to 1/20 of mainstream models like Claude Opus and GPT-5 at a throughput of 50 tokens per second [1][2] - The model has shown strong performance in programming tasks and has achieved first place in the Multi-SWE-Bench multilingual task, with a 37% improvement in task completion speed compared to its predecessor M2.1 [2] Cost Efficiency - M2.5 is designed to eliminate cost constraints for running complex agents, achieving a processing speed of 100 TPS, which is approximately double that of current mainstream models [3] - The model reduces the total token consumption for tasks, averaging 3.52 million tokens per task in SWE-Bench Verified evaluations, down from 3.72 million tokens in M2.1 [3] Programming Capabilities - M2.5 emphasizes system design capabilities in addition to code generation, demonstrating a native specification behavior that allows it to decompose functions and structures from an architect's perspective [4] - The model has been trained in over 10 programming languages and has shown a pass rate of 79.7% on the Droid platform and 76.1% on OpenCode, outperforming previous models [5] Task Handling Efficiency - In search and tool invocation, M2.5 exhibits higher decision maturity, achieving approximately 20% fewer rounds of consumption compared to previous versions while maintaining token efficiency [8] Office Applications - MiniMax has integrated industry-specific knowledge into M2.5's training, resulting in an average win rate of 59.0% in the Cowork Agent evaluation framework against mainstream models, capable of producing industry-standard reports and financial models [10] Technical Foundation - The performance improvements of M2.5 are driven by a large-scale reinforcement learning framework named Forge, which decouples the underlying training engine from the agent [14] - The engineering team has optimized asynchronous scheduling and tree-structured sample merging strategies, achieving approximately 40 times training acceleration [14] Deployment - M2.5 is fully deployed in MiniMax Agent, API, and Coding Plan, with model weights set to be open-sourced on HuggingFace for local deployment [15]