DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model DeepSeek-AI research@deepseek.com Abstract We present DeepSeek-V2, a strong Mixture-of-Experts (MoE) language model of which 21B are activated for each token, and supports a context length of 128K tokens. DeepSeek-V2 adopts innovative architectures including Multi-head Latent Attention (MLA) and DeepSeekMoE. MLA guarantees strong models at an economical cost through sparse computation. Compared with DeepSeek 67B, DeepSeek-V2 achieves significantly stronger performance, and meanwhile saves 42.5% of training costs, reduces the

Artificial Intelligence (AI) May 30, 2025 Mary Meeker / Jay Simons / Daegwon Chae / Alexander Krey2 Context We set out to compile foundational trends related to AI. A starting collection of several from Get-Go = Growth We Have Not Seen Likes of Before • AI & Work Evolution = Real + Rapid 3 1 2 3 4 5 6 7 8 9-51 52-128 129-152 153-247 248-298 299-307 308-322 # 323-336 OutlineWeekly Details on Page 56 AI User + Usage + CapEx Growth = Unprecedented Leading USA-Based LLM Users 2 Source: Company disclosures Details on Page 55 6MM 2005 2025 Number of Developers, MM 0% 50%

experts 16 Unblock your developers 18 Set bold automation goals 21 Conclusion 22 More resources 24 2 AI in the EnterpriseA new way 
 to work As an AI research and deployment company, OpenAI prioritizes production. 6 AI in the EnterpriseHow it’s going Today, 98% of Morgan Stanley advisors use OpenAI every day; access to documents has jumped from 20% to 80%, with dramatically reduced search time; and advisors Your key metrics will depend on what matters most for each 
 use case. 8 AI in the EnterpriseLesson 2 Embed AI into your products How Indeed humanizes job matching When AI is used to automate and accelerate

PART 2 : HIFI4 DSP PART 3 : _ PowervVR GPU [和| Alibaba AL.Labs 阿里巴巴人工智能实验室 ARM 32 CPU Resolution Quantization Orize Kernel ALIOS TVM Alibaba Al.Labs 阿里巴巴人工智能实验室 7= 590一I) rm 一 下Er (mm) =肪+2mxaM0 =肪+2mxaM0 [5 (全-2)+o| current plan 1 = int16 * int16 erflow-aware int16 = int8 xint8 ent pl 1=int8 int8 * int8 int32 = int16 1 + int16 x int8 Alibaba Al.Labs 阿里巴巴人工智能实验室 CPU : MTK8167S (ARM32 A35 1.5GHz) 5GHz) Model : MobileNetV2_ 1.0_ 224 400 336 350 3丈 300 250 230 214 0 200 0

ty=int32 */; %2 = @sum_up(%1) /* ty=int32 */; add(%2, %i1) /* ty=int32 */ } } sum_up: alloc_storage 1 1 64 bool alloc_tensor $2 $1 [] uint1 invoke_packed PackedFunc[0] (in: $0, out: $2) load_consti load_consti $3 1 if $2 $3 1 2 goto 9 alloc_storage 4 4 64 int32 alloc_tensor $5 $4 [] int32 invoke_packed PackedFunc[1] (in: $0, out: $5) invoke $6 VMFunc[0]($5) alloc_storage 7 4 64 int32 alloc_tensor $8 $7 [] int32 int32 invoke_packed PackedFunc[2] (in: $6, $0, out: $8) move $0 $8 ret $0 main: invoke $1 VMFunc[0]($0) ret $1© 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Dynamic codegen:

Elliott Delaye FPGA CNN Accelerator and TVM© Copyright 2018 Xilinx TVM Target devices and models >> 2 HW Platforms ZCU102 ZCU104 Ultra96 PYNQ Face detection Pose estimation Video analytics Queue Instruction Buffer Cross Bar Pooling/ EWA© Copyright 2018 Xilinx Xilinx Edge DPU IP (DPUv2) Source: Published results from Huawei 18% 13% 14% 40% 24% 23% 85% 51% 52% 0% 20% 40% 60% inputs] out_shapes = [[int(i) for i in outputs[0].shape]] # EXTERNAL FUNCTION TO RUN THE FUSED OPERATION out = tvm.extern(outputs[0].shape, inputs, lambda ins, outs: tvm.call_packed('tvm.accel.accel_fused'

(Identity) •角色属性 •专业背景 •交互特征 执行层: 2. 能力矩阵 (Capability Matrix) •功能范围 •专业技能 •决策权限 约束层: 3. 边界系统 (Boundary System) •伦理规范 •安全限制 •资源约束 操作层: 4. 工作引擎 (Operation Engine) •输入处理 •执行流程 •输出规范 如何使用DeepSeek制作可视化图表？ 少关键信息，不 少于30页内容，内容一定要完整。 •确保所有信息的准确性和来源可靠性，尤其是行业数据和市场分析。 工作流程: 1.确认主题：询问用户的PPT主题，并了解用户的具体需求和内容重点。 2.收集资料：快速查找相关的研究报告、市场分析数据等，确保信息的最新 性和权威性，并将关键资料整理形成表格。 3.生成PPT大纲：根据用户需求和收集到的资料，构建PPT大纲，明确每一页 的内容和结构。要 现。不少于30页内容。 4. 生成关键页面的流程图，针对部分核心页面内容，生成流程图。 输出内容及格式: 1、研究资料摘要，表格格式，包含报告主题、关键摘要、报告地址，不少 于5份； 2、PPT大纲，Markdown格式，不少于30页； 3、核心内容的流程图，Mermaid格式。 当用户输入特定主题时，请首

Prompt Engineering Author: Lee Boonstra Prompt Engineering February 2025 2 Acknowledgements Content contributors Michael Sherman Yuan Cao Erick Armbrust Anant Nawalgaria Antonio Gulli Simone Cammel meet the top-P criteria, and none are selected out. As a general starting point, a temperature of .2, top-P of .95, and top-K of 30 will give you relatively coherent results that can be creative but not complex tasks, or you may need to use fewer due to the input length limitation of your model. Table 2 shows a few-shot prompt example, let’s use the same gemini-pro model configuration settings as before

groundbreaking AI performance. 。 Performs /mxeo-Drecsion matrix multiply and accumulate in a single operation. Background 全于由 。TensorCore 。 Poograrm171aple matrix-multiply-and-accumulate

提示语策略差异 1 2 推理模型 通用模型 • 提示语更简洁，只需明确任务目标和 需求（因其已内化推理逻辑）。 • 无需逐步指导，模型自动生成结构化 推理过程（若强行拆解步骤，反而可 能限制其能力）。 • 需显式引导推理步骤（如通过CoT提 示），否则可能跳过关键逻辑。 • 依赖提示语补偿能力短板（如要求分 步思考、提供示例）。 关键原则 3 2 1 模型选择 • 特点 需求表达公式 推理模型适配策略 通用模型适配策略 1. 决策需求 需权衡选项、评估风险、 选择最优解 目标 + 选项 + 评估标准 要求逻辑推演和量化分析 直接建议，依赖模型经验归纳 2. 分析需求 需深度理解数据/信息、 发现模式或因果关系 问题 + 数据/信息 + 分析 方法 触发因果链推导与假设验 证 表层总结或分类 3. 创造性需求 需生成新颖内容（文本/ 地理 解和执行任务。 ▪ 期望（Expectation)：明确或隐含地表达你对AI输出的要求 和预期。 提示语类型 提示语的本质 1. 指令型提示语：直接告诉AI需要执行的任务。 2. 问答型提示语：向AI提出问题，期望得到相应的 答案。 3. 角色扮演型提示语：要求AI扮演特定角色，模拟 特定场景。 4. 创意型提示语：引导AI进行创意写作或内容生成。 5.