经济观察网12月28日，恒生电子（600570）联合创始人刘曙峰在参加中国财富管理50人论坛2025年会时 表示，2025年大模型的产业应用取得实质性突破。以金融为例，场景集中体现交互入口、文档信息处 理、客户服务等场景下，同时，大模型在代码生成环节的生产有效性也得到了印证。在智能体Agent的 发展领域，有望在未来一年取得实质性的进展。在精准计算领域，传统小模型对结构化数据的处理能力 不能也无需被替代，大小模型的混合使用可以是一种有效的方式。与此同时，大模型的幻觉问题无法从 根本上消除，需要探索有效的边界并接受与大模型幻觉共存，在金融领域，这意味着AI在风险管理， 投资决策等高价值部位的有效使用。 刘曙峰表示，大模型金融应用仍处于初级阶段，限制发展速度的主要因素包括私有化部署的有效性，业 务的合规风险以及预算的约束。对很多机构而言，观望不失为一种有限策略，不必焦虑。从长期战略的 角度看，底层技术的进步最终会改变商业模式和产业范式的基础架构，最新的观察是数据中台和AI中 台的融合，以及所谓"本体"的业务逻辑模型，行业know-how的深度和抽象能力仍然是核心竞争力的来 源。 ...

Investment Rating - The report rates the industry as "Outperform" [1] Core Insights - The release of Alibaba's Qwen3 confirms that leading AI companies in China are at the forefront of global technology, with open-source models expected to significantly boost the AI industry [2][9] - Qwen3 achieved a new high in the BFCL evaluation, indicating strong support for the upcoming AI Agent era [2][12] - The report maintains a positive outlook on the computer sector and suggests monitoring specific companies such as Guangzhou Sie Consulting, ArcSoft Corporation, Hygon Information Technology Co., Ltd., and others [2][9] Summary by Sections Qwen3 Model Performance - Alibaba launched Qwen3, the world's strongest open-source model, with the flagship model Qwen3-235B-A22B surpassing top competitors like DeepSeek-R1 and OpenAI's models [10][12] - Qwen3's dataset has expanded to approximately 36 trillion tokens, nearly double that of its predecessor Qwen2.5, covering 119 languages [11] - Qwen3 supports two thinking modes: a thoughtful mode for complex problems and a quick mode for simpler queries, enhancing its operational efficiency [11] Agent Capabilities - Qwen3 excels in the Agent domain, achieving a score of 70.8 in the BFCL evaluation, surpassing other leading models [12] - The introduction of Qwen-Agent simplifies the integration of tools, enhancing the model's capabilities in real-world applications [12] Investment Recommendations - The report highlights several companies to watch, including 合合信息 (Hehe Information), 赛意信息 (Saiyi Information), 鼎捷数智 (Dingjie Smart), and others, with detailed earnings forecasts provided [6][9]

Qwen3 在推理、指令遵循、工具调用、多语言能力等方面均大幅增强。在官方的测评中，Qwen3 创下所有国产模型及全球开源模型的性能新高：在奥 数水平的 AIME25 测评中，Qwen3 斩获 81.5 分，刷新开源纪录；在考察代码能力的 LiveCodeBench 评测中，Qwen3 突破 70 分大关，表现甚至超过 Grok3；在评估模型人类偏好对齐的 ArenaHard 测评中，Qwen3 以 95.6 分超越 OpenAI-o1 及 DeepSeek-R1。 | | Qwen3-235B-A22B | Qwen3-32B | OpenAl-o1 | Deepseek-R1 | Grok 3 Beta | Gemini2.5-Pro | Open Al-o 3-mini | | --- | --- | --- | --- | --- | --- | --- | --- | | | MoE | Dense | 2024-12-17 | | Think | | Medium | | ArenaHard | 95.6 | 93.8 | 92.1 | 93.2 | - | 96.4 | 89.0 | | AIM ...