今天大家都在谈MS的这篇DeepSeek R2分析的报告，提前曝光了R2的性能和参数，我们简单总结一 下这个报告的核心内容： DeepSeek R2 使用了多达 1.2 万亿个参数，采用了新颖的架构，实现了运行成本的显著降低。其采用 混合专家混合（MoE）架构，有 780 亿个活跃参数。 并且R2 使用华为的 Ascend 910B 芯片进行训练，而非 NVIDIA 的芯片。 R2 增强了多语言覆盖能 力，能流畅处理非英语语言；扩展了强化学习，利用更大的数据集，使模型能够进行更具逻辑性和 更像人类的推理；增加了多模态功能，能够处理文本、图像、语音和视频数据；实现了推理时的缩 放，通过采用通用奖励模型（GRM），在推理过程中增加计算资源，从而提高了输出质量。 R2 具有高成本效益，输入成本为每百万代币 0.07 美元，输出成本为每百万代币 0.27 美元，而 R1 的 输入成本为 0.15-0.16 美元，输出成本为 2.19 美元。 由于这篇报告讲的人已经很多了，我们就不赘述了，而且报告也放到了星球中，有兴趣的朋友可以 到星球中看原文。 今天这篇文章来看另一篇AI的分析，Artificial Analysis ...

Core Insights - The report on DeepSeek R2 highlights its significant advancements in performance and cost efficiency, utilizing a novel architecture with 1.2 trillion parameters and a mixture of experts (MoE) framework [1] - The report from Artificial Analysis outlines six major trends in the AI sector expected by early 2025, focusing on advancements in intelligence, efficiency, and multimodal capabilities [2] Group 1: AI Progress - The AI industry continues to make strides in model intelligence, cost efficiency, and speed, with leading labs like OpenAI, Google, and xAI at the forefront [3] - OpenAI's o4-mini and o3 models lead in intelligence, followed by Google's Gemini 2.5 Pro and xAI's Grok 3, indicating a competitive landscape with rapid innovation [3] - OpenAI and Google maintain a competitive edge through vertical integration in the AI value chain, while smaller players focus on specific modalities [3] Group 2: Rise of Chinese AI - Chinese AI labs, such as DeepSeek and Alibaba, have made significant progress in open-weight models, narrowing the gap with U.S. labs and enhancing China's influence in the open AI ecosystem [4] Group 3: Reasoning Models - Reasoning models that generate intermediate tokens before answering have significantly improved intelligence levels, outperforming non-reasoning models in various assessments [5] - Google’s Gemini 2.5 Pro exemplifies this advancement by correctly answering complex problems, while non-reasoning models prioritize speed and cost [5] Group 4: AI Agents - AI systems are increasingly capable of autonomously completing end-to-end tasks by chaining requests from multiple large language models (LLMs), enhancing their practicality [6] Group 5: Efficiency and MoE - The report emphasizes that advancements in small model intelligence, reasoning efficiency, and next-generation hardware have led to a significant reduction in inference costs [7] - MoE models activate only a portion of parameters during inference, contributing to improved efficiency and accessibility of high-performance AI [7] Group 6: Multimodal AI - Multimodal AI has made substantial progress, with advancements in image generation, video generation, and speech processing [8][9] - OpenAI's GPT-40 sets a new standard in image generation quality, while Google’s Veo 2 surpasses OpenAI's Sora in video generation [8] - Speech-to-text and text-to-speech models have also improved, with OpenAI and ElevenLabs leading in accuracy [9] Group 7: Open-Weight Models and Competitive Landscape - Open-weight models from Alibaba, DeepSeek, Meta, and NVIDIA have significantly closed the intelligence gap with proprietary models, although OpenAI's o4-mini and Google's Gemini 2.5 Pro still hold slight advantages [14] - The AI landscape is becoming increasingly crowded, with competition among U.S. labs and companies like NVIDIA, DeepSeek, and Alibaba intensifying [14]

长按即可参与 到现在为止，大多数人可能都听说过中国人工智能初创公司 DeepSeek，因为它当时几乎在所有平台 都爆红。但我们认为因为 DeepSeek（ DEEPSEEK ）而抛售英伟达股票是不合理的，至少没有达 到那种程度。 但让我们仔细分析一下市场为何会有如此反应，这有助于我们理解如果Deepseek R2人工智能模型 的发布是否会重演这一幕。在R1发布之前，人们普遍认为中国在人工智能领域落后多年，没有机会迎 头赶上。 谷歌前首席执行官埃里克·施密特 (Eric Schmidt)曾在 2024 年 5 月表示，美国在人工智能领域领先 中国 2-3 年，原因有三：由于芯片禁令，中国更难获得英伟达芯片；获取培训材料更困难，互联 网上的信息更多是英文；最后，投资较少。领先两三年意味着中国现在的水平与 ChatGPT 的第 一个版本相同。然而，事实并非如此。 顺便说一句，DeepSeek 之前也发布过一些模型，但由于性能不佳，它们从未声名鹊起，也鲜有人使 用。然而，R1 的发布改变了一切，并表明使用更少、更老旧的 GPU 也能开发出最先进的模型。开 发人员运用了多种优化策略来实现这一点。然而，黄仁勋表示，在他看 ...