Quantitative Models and Construction Methods Model: DeepSeek-R1 - **Model Construction Idea**: The DeepSeek-R1 model aims to innovate in AI technology by reducing dependency on high-end imported GPUs and enhancing cost-effectiveness and performance in global markets[5][12][30] - **Model Construction Process**: - The model is based on the DeepSeek-V3 architecture and applies reinforcement learning techniques during the post-training phase to significantly improve inference capabilities with minimal labeled data[33] - The model's performance in tasks such as mathematics, coding, and natural language inference is on par with OpenAI's o1 official version[33] - The team also introduced six distilled small models using knowledge distillation techniques, with the 32B and 70B versions surpassing OpenAI o1-mini in several capabilities[34] - The model's training cost was $5.576 million, only 1/10th of GPT-4o's training cost, and its API call cost is 1/30th of OpenAI's similar services[38] - **Formula**: $$ \text{SUE} = \frac{\text{Single Quarter Net Profit} - \text{Expected Net Profit}}{\text{Standard Deviation of Net Profit YoY Change over the Past 8 Quarters}} $$ where Expected Net Profit = Last Year's Same Quarter Actual Net Profit + Average YoY Change in Net Profit over the Past 8 Quarters[55] - **Model Evaluation**: The model is highly cost-effective and adaptable to different application environments, breaking the traditional AI industry's reliance on "stacking computing power and capital"[38][43] Model Backtesting Results - **DeepSeek-R1 Model**: - **AIME pass@1**: 9.3 - **AIME cons@64**: 13.4 - **MATH-500 pass@1**: 74.6 - **GPQA Diamond pass@1**: 49.9 - **LiveCodeBench pass@1**: 32.9 - **CodeForces rating**: 759.0[36] Quantitative Factors and Construction Methods Factor: Standardized Unexpected Earnings (SUE) - **Factor Construction Idea**: SUE is used to measure the growth potential and latest marginal changes in the prosperity of the industry and individual stocks[57] - **Factor Construction Process**: - SUE is calculated as: $$ \text{SUE} = \frac{\text{Single Quarter Net Profit} - \text{Expected Net Profit}}{\text{Standard Deviation of Net Profit YoY Change over the Past 8 Quarters}} $$ where Expected Net Profit = Last Year's Same Quarter Actual Net Profit + Average YoY Change in Net Profit over the Past 8 Quarters[55] - **Factor Evaluation**: SUE effectively measures future earnings growth and the latest marginal changes in prosperity, representing the future trend changes in the industry[57] Factor Backtesting Results - **SUE Factor**: - **2022**: -29.8% - **2023**: 15.9% - **2024**: 20.1% - **2025 YTD**: 11.0%[65]

Group 1: Technology Updates - Google has updated its Gemini 2.5 model family, introducing features like a "thinking" mechanism and multi-modal input capabilities, with a pricing standard of $0.3 per million tokens for input [2] - MiniMax launched the MiniMax-M1 model, the world's first open-source large-scale hybrid architecture inference model, supporting 1 million context inputs and 80,000 token outputs, with reinforcement learning costs reduced to $530,000 [3] - The new open-source model Kimi-Dev-72B from 月之暗面 achieved the highest score in the SWE-bench Verified programming benchmark, scoring 60.4% with only 72 billion parameters [4] Group 2: Automotive and Delivery Services - 鸿蒙智行 achieved a weekly delivery of 11,600 units, maintaining the top position among new force brands for four consecutive weeks, with flagship model 问界M8 surpassing 5,000 units in weekly deliveries [5] - 京东's CEO reported that the company has over 120,000 full-time delivery riders, with daily order volume exceeding 25 million and an average monthly income of 13,000 yuan for riders in major cities [7] Group 3: Strategic Partnerships and Collaborations - 传音控股 signed a strategic cooperation agreement with Indonesian telecom operator IOH to enhance 5G terminal penetration and improve user conversion rates [12] - 德马科技 and 上海智元新创 signed a strategic cooperation agreement to explore innovative applications of humanoid intelligent robots in logistics [11] Group 4: Financial and Market Activities - 兆芯集成's IPO application has been accepted, aiming to raise 4.169 billion yuan for new generation processor projects [16] - 曹操出行 announced its IPO plans, targeting to raise 1.853 billion HKD with a share price of 41.94 HKD [17] - 京东方A plans to acquire a 30% stake in 咸阳彩虹光电科技有限公司, with a base price of 4.849 billion yuan [18] Group 5: Industry Developments - The domestic silicon carbide IDM company 芯聚能半导体 has successfully scaled its SiC chip production for automotive applications [14] - The memory market is experiencing significant price increases for DDR4 memory, although market transactions are showing signs of weakness [15]

文 / 于钟海 , 魏鹳霏 , 肖楷 , 赵丽萍 中金研究 以MiniMax V-Triune新框架成果为例，推理感知统一框架在可拓展性、泛化性初步验证。 V-Triune以三层组件架构实现视觉推理和感知任务统一至强化学 习框架：1）多模态样本数据格式化；2）验证器奖励计算，采用异步客户端-服务器架构，奖励计算和主训练循环解耦；3）数据源级指标监控，便于溯源 和提升稳定性。结合动态IoU奖励机制、冻结ViT参数等工程优化，Orsta系列模型（32B参数）在MEGA-Bench Core基准测试中实现了最高14.1%的性能提 升。 多模态推理助力智能驾驶能力升阶。 在智能驾驶场景，多模态推理是增强道路交通标志识别判断能力、提升复杂场景泛化性的重要途径，正成为头部智 能驾驶企业算法演进的重点。2025年5月30日，蔚来世界模型NVM首个版本正式开启推送，具备全量理解、想象重构和推理能力，能够对实时环境多模信 息进行理解和推演，在选择最优ETC车道通行、停车场自主寻路等场景的性能提升显著。此外，理想自研的VLA大模型亦具备思维链推理能力，以多模态 推理模拟人类驾驶员的思维运作方式。 图表1：MiniMax多模态RL ...