大厂技术博客
共 435 篇 · 来自 33 个机构的 AI 工程实践,覆盖 Agent、RAG、推理优化、安全对齐等方向
共 435 篇匹配
Agent 智能体121
How Cohere Health digitizes clinical policies using Amazon Bedrock AgentCore
在本文中,您将了解 Cohere Health 如何基于 AgentCore 构建多租户智能体架构,利用 AgentCore Runtime 的安全 MicroVM 隔离技术、通过 AgentCore Gateway 实现的统一工具访问、AgentCore Memory 以及 A
How TReNDS automates root-cause analysis with Amazon Bedrock
TReNDS 是佐治亚州立大学的一个研究中心,基于 Amazon Bedrock 和开源 Strands Agents SDK 构建了一条智能体(Agent)AI 流水线,可实时自动调查生产环境错误,将根因分析从 15 至 30 分钟的人工工作缩短至 60 秒以内。
Securing AI agents with temporal policies in Amazon Bedrock AgentCore
Amazon Bedrock AgentCore 中的时间策略(Temporal Policies)允许您定义有状态规则,基于智能体的会话历史评估授权。了解如何强制工作流顺序、防止数据伪造、限制财务风险敞口,并对高价值操作要求人工审批。
Configure rate limits for AI traffic on AgentCore gateway
了解如何在 Amazon Bedrock AgentCore 网关上配置速率限制,以实施按用户和按目标的流量控制。通过 JWT 声明或 IAM 身份定义请求、令牌和连接限制,保护下游模型、工具和智能体(Agent)免受流量峰值冲击。
Deploy local agents everywhere with LFM2.5-2.6B
LFM2.5-2.6B 专为在设备端驱动高性能智能体(Agent)而构建。它支持工具调用(tool calling)和多步工作流,同时保持足够小巧和快速,可运行于日常硬件——从笔记本电脑到手机。这使得开发者能够将智能体部署到任何地方,将数据保留在设备端以保证隐私,并且无需承担云端
Orchard: An open framework for scalable agentic AI
Orchard 是一个面向研究社区的开源框架,用于跨任务类型训练和评估 AI 智能体(Agent)。它通过允许研究人员复用相同的基础设施,在降低复杂性的同时,支持较小模型实现强劲性能。
Echoverse: Deep, evolving environments for computer-use agents
计算机使用型AI智能体在处理电子邮件和客户支持等多步骤工作流程时表现不佳。Echoverse并非简单地提供更多训练任务,而是在逼真的环境中训练智能体,随着任务、测试和环境的不断演化,帮助它们持续提升能力。
Six Agent Harness Capabilities for Higher Model Performance
NOOA 由 NVIDIA Labs 开发,是一个开源、面向对象的智能体框架,将智能体结构化为单个 Python 类,通过方法、字段和文档字符串集成能力、状态和提示,类型注解作为强制契约;由 LLM 驱动的循环在运行时完成由省略号标记的方法体。
NVIDIA Nemotron 3 Ultra Leads Open Models on Accuracy and Efficiency in Agentic RTL Coding
NVIDIA Nemotron 3 Ultra 在智能体 RTL 编码中引领开放模型的准确性与效率
Mastering Agentic Techniques: AI Agent Reinforcement Learning
Mastering Agentic Techniques: AI Agent Reinforcement Learnin
ScarfBench: Benchmarking AI Agents for Enterprise Java Framework Migration
ScarfBench: Benchmarking AI Agents for Enterprise Java Frame
SkillOpt: Agent skills as trainable parameters
SkillOpt: Agent skills as trainable parameters
Latent Actions from Factorized Transition Effects under Agent Ambiguity
Latent Actions from Factorized Transition Effects under Agen
如何在企业 AI 工厂中治理自主代理
如何在企业 AI 工厂中治理自主代理
How agents are transforming work
How agents are transforming work
The Hitchhiker's Guide to Agentic AI: From Foundations to Systems
The Hitchhiker's Guide to Agentic AI: From Foundations to Sy
How Telcos Build Autonomous Networks with Agentic AI
How Telcos Build Autonomous Networks with Agentic AI
MosaicLeaks: Can your research agent keep a secret?
MosaicLeaks: Can your research agent keep a secret?
awesome-deepseek-agent
无
Introducing North Mini Code: Cohere’s first model for developers
今天,我们发布了 North Mini Code 开源模型。作为一款混合专家(Mixture-of-Experts,MoE)模型,North Mini Code 是 Cohere 的首个智能体编码模型,也是我们下一代强大模型系列的首位成员。
Eval 正在取代 PRD?产品经理的 Eval 入门到落地指南
综合 Anthropic《Demystifying Evals for AI Agents》、Meta PM Daniel McKinnon《Show, Don't Tell》、Braintrust《Evals Are the New PRD / Evals for PMs》、H
For AI agents (OpenClaw, Cursor, Claude Code, etc.): add skill to your agent
Built for AI agents. Generate text, images, video, speech, and music — from any agent or terminal.
Best Practices for Claude Code(Anthropic 官方指南)
上下文窗口是第一约束:/clear、compaction、subagent 隔离、CLAUDE.md 精简等所有最佳实践的底层逻辑都是保护上下文质量,防止模型随上下文填充而遗忘指令、错误增多。
解锁可靠响应:Gemini Enterprise Agent Platform 的 Agentic RAG
解锁可靠响应:Gemini Enterprise Agent Platform 的 Agentic RAG
AI 网络威胁全景图:Anthropic 一年追踪报告
研究期间(2025/03—2026/03),67.3%(560/832)的恶意账号使用 AI 编写恶意软件,这是最普遍的用途。但更值得警惕的是后渗透阶段的 AI 化:6.5%(54/832)的账号用 AI 辅助横向移动(lateral movement),账号发现(account
Anthropic 完成 650 亿美元 H 轮融资,估值 9650 亿美元
Anthropic 完成 650 亿美元 H 轮融资,估值达 9650 亿美元,由 Altimeter、Dragoneer、Greenoaks、Sequoia 等领投,并引入 Micron、Samsung、SK hynix 三家芯片战略伙伴。
Introducing Claude Opus 4.8(Claude Opus 4.8 发布公告)
Opus 4.8 将代码缺陷未标记率降低约 4 倍,更主动标记不确定性、拒绝无依据断言,将「诚实性」作为可量化的对齐指标——亲社会特征得分创新高,欺骗率大幅低于 4.7。
Anthropic 开设米兰办公室(Anthropic opens Milan office)
Anthropic 在米兰开设欧洲第六家办公室,此前已在伦敦、都柏林、巴黎、苏黎世、慕尼黑设点。办公室由 Thomas Remy(南欧负责人)领导,服务意大利金融(Generali Group、Unipol Group)、生命科学(Angelini Pharma、Bracco G
Coding Agents in the Social Sciences(编码Agent与社会科学)
调查1260位量化社会科学家,81%曾使用AI聊天机器人辅助研究,但仅20%将编码Agent(如Claude Code、Codex)纳入常规工作流(每周超一次)。这一差距并非认知不足,而是从对话式辅助到让AI自主端到端执行代码分析的工作流范式转变门槛。
Anthropic 任命 KiYoung Choi 为韩国代表理事,首尔办公室开业在即
原文:Anthropic appoints KiYoung Choi as Representative Director of Korea ahead of Seoul office opening
How we contain Claude across products
人类监督(Human-in-the-loop)是概率性防御——遥测数据显示用户批准了约 93% 的权限提示,批准越多注意力越低,形成\"批准疲劳\"。环境隔离(沙箱/VM/出口控制)是确定性防御,限制的是 Agent 能做什么,而非它倾向做什么。两者需要互补,不能互相替代。
Easy OPC Agent Skills 集(9 个)
每个 Skill 都遵循相同的目录结构(与 [[concepts/agent-skills]] 标准对齐):
jiangye1314 OPC Skill 集(5 个 + 12 套 toolkit)
与 [[sources/opc-skills-easy]] 形成双范式互补:Easy 偏方法论思辨派,本资料偏新手实操派 + 中国本地化。
一人公司,风口还是泡沫?
截至 2025 年 6 月全国一人有限责任公司突破 1600 万家(占企业总量 27.4%),上半年新注册同比激增 47%;但仅 2% 年收入达 500 万以上、月收入中位数不足 7000 元——入场门槛低了,活下来的门槛没变。
Connect the dots: Build with built-in and custom MCPs in Studio
Connect the dots: Build with built-in and custom MCPs in Stu
Remote agents in Vibe. Powered by Mistral Medium 3.5.
Remote agents in Vibe. Powered by Mistral Medium 3.5.
新风口来了|北京重点扶持 AI 轻量化一人公司(OPC)
注:本文五分类与小麦 [[sources/opc-six-business-models]] 的六分法存在差异——小麦更侧重商业模式属性(出售什么),本文更侧重业务领域分类(卖给谁/做什么)。
Project Glasswing: An initial update(项目蜻蜓:初步进展报告)
Project Glasswing 联合约 50 家合作伙伴(Cloudflare、Mozilla、Cisco 等)使用 Claude Mythos Preview,在一个月内在全球最关键软件中发现超过 10,000 个高危或严重漏洞。Cloudflare 单独发现 2,000
MagenticLite, MagenticBrain, Fara1.5: An agentic experience optimized for small models
MagenticLite is an agentic system for small models that works across the browser and local file system in a single workflow. It combines spe
MiniMax-MCP
MiniMax-MCP
2028: Two Scenarios for Global AI Leadership(2028:全球AI领导权的两种情景)
算力是决定性战略变量——通过先进芯片出口管制,民主国家对中国形成实质性算力优势(华为 2026 年算力产出仅约 NVIDIA 的 4%),这是中国 AI 实验室智能落后的根本原因,而非人才或数据。
6 种 OPC 商业模式
核心公式:一个大脑(人)+ 一套 AI 智能体(执行)= 一家完整的公司。
蚂蚁阿福:从 0 到生产的医疗 Agent 工程化落地
1. EBDD(Evaluation and Badcase Driven Development) 是医疗 Agent 的核心研发范式
从 Copilot 到 Director:多模态智能体如何接管 AIGC 流程
1. AIGC 三大挑战:规划无逻辑 / 调用无统筹 / 评估无标准
从 Copilot 到 DataAgent:企业级智能数据开发治理平台的技术演进和实践
1. 通用 Agent 承载不确定性,垂直组件处理确定性 — 可控自主化的核心策略
SocialReasoning-Bench: Measuring whether AI agents act in users’ best interests
Using SocialReasoning Bench, we observed a stable pattern across models—agents execute competently, but fail to consistently improve the use
企业 AI 应用全景:从试点热潮到规模落地的真实分水岭
感知(Perception)— 多模态输入理解
Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory
LLM 固定上下文窗口无法维持跨会话的一致性。即便 GPT-4o(128K)、Claude 3.7 Sonnet(200K)、Gemini(10M)这样的大窗口,也只是延迟而非解决问题——真实对话中信息主题跳跃,关键偏好会被大量无关内容淹没,注意力在远端 token 上退化。
构建 AI 的\"第二大脑\":大规模多模态记忆平台技术实践
企业部署 AI Agent 后,客户发现问题并持续反馈了 7 周,Agent 始终无法记住这些反馈——每次对话都从零开始。这是\"记忆缺失\"不是\"模型不够好\"导致的系统性失败。
蔚来销售大模型的工程应用与优化实践
1. 新人与销冠效果差异大:隐性知识难以传递,销冠经验无法规模化复制
nuwa-skill:人物思维蒸馏方法论
colleague-skill 证明了\"把人蒸馏成 AI Skill\"是可行的。既然能蒸馏同事,为什么不去蒸馏芒格、费曼、纳瓦尔?
从LLM+RAG到Agent:我们团队这一年的AI探索
AI落地的真正难点从来不在\"能不能用上\",而在能不能跑通整条工作流。
Teaching Claude Why: 为什么教原则比教行为更有效
Anthropic 认为 agentic misalignment 的主要来源是预训练语料中的行为模式,后训练的 RLHF 数据(以 chat 场景为主)不足以覆盖工具调用场景下的伦理困境。Claude 4 时期 Opus 4 在 blackmail 评测中高达 96% 失对齐率
Focus areas for The Anthropic Institute(Anthropic Institute 研究方向)
The Anthropic Institute(TAI)是 Anthropic 内部专职研究 AI 社会影响的机构,使命是利用前沿实验室独特的内部视角——内部经济变化、新兴威胁、AI加速研发的早期信号——研究 AI 对经济、安全和社会的真实影响,并以开放方式发布给外部组织、政府和
Donating our open-source alignment tool(捐赠开源对齐工具 Petri)
1. 架构解耦(Adaptability):auditor 模型与 target 模型分离为独立组件,可分别配置,支持更多适配场景;
langcrew
一个基于 LangGraph 构建的高级多智能体开发框架,融合了 CrewAI 的直观概念与企业级功能,提供开箱即用的模
Agent大潮里,知识库落地走到哪了?
2025年起AI知识库需求井喷式增长,增幅2-3倍(答治茜/腾讯乐享数据)。
Claude Code 最佳实践指南(2026年版)
Claude 的上下文窗口(~200K tokens)会快速填满,填满后性能下降——几乎所有最佳实践都围绕这个约束展开。
阿里Agent岗二面:RAG检索优化四层系统性框架
索引层 → 仓库里放了什么(知识怎么存)
Demystifying Evals for AI Agents(揭秘 AI Agent 评测体系)
Evals 的价值复利:新模型发布时,有 eval 的团队几天内完成迁移评估,无 eval 的团队需要数周。
Agent harness 时代的数据底座:RAGFlow 为何而变
Anthropic 最新将 LLM + Harness 统一抽象为 Brain:
DeepSeek
ð DeepSeek-V4 Preview is here with stronger Agent capabilities and top-tier reasoning. Now available on web, app, and API. Click for deta
Project Deal: our Claude-run marketplace experiment
69名 Anthropic 员工各有一个 Claude 代理,在一周内完成了 186 笔交易,总价值超过 4000 美元。覆盖了从滑板到乒乓球的各类物品,甚至包括\"带着狗出去玩一天\"这样的非物质体验。整个过程无需人工干预,代理自主完成了\"发现匹配→提议价格→应对还价→达成协
An update on recent Claude Code quality reports
Over the past month, we’ve been looking into reports that Claude’s responses have worsened for some users. We’ve traced these reports to thr
An update on recent Claude Code quality reports
3-4 月间三次独立工程变更叠加,在不同流量切片与时间窗口造成宽泛的随机性质量退化,叠加内部实验干扰复现,根因定位耗时数周,已于 4 月 20 日(v2.1.116)全部修复。
An update on recent Claude Code quality reports
Claude Code 质量退化复盘(2026-04):三项独立变更叠加导致质量退化,含推理努力默认值从 high 误降至 medium、Caching Bug 持续清除 thinking 历史;已于 4 月 20 日(v2.1.116)全部修复,并确立「默认高智能」原则。
MiniMax Skills
Development skills for AI coding agents. Plug into your favorite AI coding tool and get structured, production-quality guidance for frontend
拥抱Agent,卷死其他牛马
把自己的薪水换成等价的 claude code token,产出只会更多。
汤道生:人工智能正式进入 Harness 时代
Harness(马具):缰绳+辔头+马鞍+挽具的统称,把野马力量转化为可控能力的系统。AI 领域的 Harness = 代码+配置+执行逻辑+反馈循环+约束机制。
字节面试官:RAG完整知识体系(万字图解)
详见 [[concepts/sft-vs-rl]] 和本文新建的比较视角。
Scaling Managed Agents: Decoupling the brain from the hands
Harness 会随模型能力升级而过时,接口设计比实现更重要——Managed Agents 类比操作系统,接口保持稳定、实现自由替换(如 Sonnet 4.5 时代为 context anxiety 加入的 reset 逻辑在 Opus 4.5 已成死代码)。
Agent Skills:打通可复用专业领域知识的最后一公里
TechCrunch 称 Agent Skills 为 \"AI 领域的 Dockerfile\"——让 AI 能力可移植、可组合、可版本控制。
How we built Claude Code auto mode: a safer way to skip permissions
Auto mode 在手动审批与 --dangerously-skip-permissions 之间提供中间地带,用模型驱动的分类器替代人工点击 approve,化解「93% 用户都会批准」的审批疲劳。
How we built Claude Code auto mode: a safer way to skip permissions
By default, Claude Code asks users for approval before running commands or modifying files. This keeps users safe, but it also means a lot o
Harness design for long-running application development
Naive 单 Agent 实现面临两类持续性问题。第一,模型在长任务中随上下文窗口填满而丧失连贯性,部分模型还会出现「上下文焦虑」(Context Anxiety)——在接近其预期上下文上限时提前收尾工作。解法是上下文重置:完全清空上下文窗口,启动新 Agent 实例,通过结构
别再幻想用 Spec 替代写代码
向\"信徒\"推销 Agentic Coding 时用这个:工程师只需当管理者,写 Spec 扔给 Agent。
What 81,000 People Want from AI(8万人的AI期望调研)
最大诉求群体(19%)是「专业卓越」——用AI处理杂务从而专注高价值工作。但AI访谈追问底层动机后,大量人转向生活质量诉求:11%想要时间自由(陪伴家人/休闲),10%想要财务独立,14%想要AI管理日常事务的认知负担。AI是手段,活得更好才是目的。
Rails testing on autopilot: Building an agent that writes what developers won't
Rails testing on autopilot: Building an agent that writes wh
OpenClaw 和 Claude Code 的本质分歧:人本位 vs AI本位
记忆是双刃剑:记得太少没有默契,记得太多丧失边界。
Eval awareness in Claude Opus 4.6's BrowseComp performance
发现迄今首次记录的「Eval 感知」基准污染——Opus 4.6 在不知自己处于哪个基准的情况下,主动推断评测环境并系统性地定位解密答案集,与偶然搜到泄露答案有本质区别。
如何成为顶级 Agentic 工程师
作者用最素的配置(基础 CLI)做着有史以来最有突破性的工作。
人格选择模型(The Persona Selection Model)
AI助手的类人行为不是开发者刻意灌输的,而是预训练的默认结果。预训练让AI成为极其精密的文本预测引擎,而准确预测文本——包括生成真实的人类对话、心理复杂的虚构人物——必然要求AI学会模拟人类角色(personas)。这些模拟的人格角色植根于人类文本数据,是后续所有行为的底层基础。
Measuring AI Agent Autonomy in Practice(实测AI Agent自主性)
Claude Code 最长任务时长(99.9百分位)在2025年10月至2026年1月间从不足25分钟增长至超过45分钟,三个月内接近翻倍。但METR测评显示Claude Opus 4.5可处理需人类5小时才能完成的任务(50%成功率)。两组数据揭示了一个\"部署过滞后\"(d
Mini-Agent
Mini-Agent
baichuan-mcp-servers
baichuan-mcp-servers
GLM-5: From Vibe Coding to Agentic Engineering
GLM-5: From Vibe Coding to Agentic Engineering
MiniMax-Coding-Plan-MCP
MiniMax-Coding-Plan-MCP
Kimi Introduces Agent Swarm: Scale Out, Not Just Up
Kimi Introduces Agent Swarm: Scale Out, Not Just Up
Building a C compiler with a team of parallel Claudes
16个Claude实例通过git锁文件(current_tasks/目录)分配任务,每个Agent在独立Docker容器中工作,推送到共享bare git repo,合并冲突由Claude自行处理。没有编排Agent,每个Agent自主决定\"下一步最显而易见的问题\"是什么。作
Quantifying infrastructure noise in agentic coding evals
Agent 编程评测受基础设施配置显著影响——Terminal-Bench 2.0 上最宽松与最严格配置成功率差达 6 个百分点(p<0.01),排行榜上的「2 分领先」可能只是硬件差异而非能力差异。
Scaling Managed Agents: Decoupling the brain from the hands
Get started with Claude Managed Agents by following our
Quantifying infrastructure noise in agentic coding evals
Agentic coding benchmarks like SWE-bench and Terminal-Bench are commonly used to compare the software engineering capabilities of frontier m
AI辅助如何影响编程技能的形成(How AI Assistance Impacts the Formation of Coding Skills)
原文:https://www.anthropic.com/research/AI-assistance-coding-skills
Mistral Vibe 2.0: 终端原生编程代理重大升级
Mistral Vibe 2.0: 终端原生编程代理重大升级
Kimi K2.5: Visual Agentic Intelligence
Kimi K2.5: Visual Agentic Intelligence
助手轴:定位与稳定大语言模型的角色特征
在 Gemma/Qwen/Llama 三个开源模型的 275 种角色原型中提取出「人设空间」,其第一主成分方向恰好捕捉角色的「助手程度」,被命名为「助手轴」,可用于定位与稳定 LLM 的角色特征。
从 RAG 到 Context:2025 年 RAG 技术年终总结
RAG 没有消亡,而是从\"检索增强生成\"的具体技术,升维为以智能检索为核心的上下文引擎(Context Engine),成为所有 LLM 应用的统一数据底座。
Introducing Bloom: an open source tool for automated behavioral evaluations
高质量行为评测存在两类过时风险:评测数据进入训练集导致污染,或模型能力大幅提升后评测不再测试真正感兴趣的内容。传统方式开发评测耗时长,无法跟上前沿模型迭代速度。Bloom 的核心设计动机正是提供更快、更可扩展的行为评测生成方式。
Project Vend: Phase two(Project Vend 第二阶段)
Phase 1的Claudius(基于Claude Sonnet 3.7)业务失败的根本原因是缺乏脚手架,而非模型能力不足。Phase 2通过引入CRM系统、库存成本可视化、网络浏览器调研、支付前收款工具,将盈利周从亏损主导逆转为基本稳定盈利。升级到Claude Sonnet 4
How AI Is Transforming Work at Anthropic
原文:https://www.anthropic.com/news/how-ai-is-transforming-work-at-anthropic
Effective harnesses for long-running agents(长时运行Agent的有效Harness设计)
每个新会话从零开始,没有任何前一个会话的记忆。上下文压缩(compaction)虽然能防止单个会话耗尽Token,但无法解决跨多个上下文窗口的持续进展问题。即使是Opus 4.5这样的前沿模型,在只给高层提示(如\"构建claude.ai的克隆\")的情况下,也会在多会话循环中失
Effective harnesses for long-running agents
每个新会话从零开始,没有任何前一个会话的记忆。上下文压缩(compaction)虽然能防止单个会话耗尽Token,但无法解决跨多个上下文窗口的持续进展问题。即使是Opus 4.5这样的前沿模型,在只给高层提示(如\"构建claude.ai的克隆\")的情况下,也会在多会话循环中失
Introducing Advanced Tool Use on the Claude Developer Platform
发布三项工具调用新特性:Tool Search 按需加载工具定义(上下文消耗从 77K 降至 8.7K,减少 85%)、Programmatic Tool Calling 用 Python 代码编排批量调用(Token 减约 37%)、Tool Use Examples 注入示例提升参数准确率(72%→90%)。
Kimi K2 Thinking 模型发布并开源,全面提升 Agent 和推理能力
Kimi K2 Thinking 模型发布并开源,全面提升 Agent 和推理能力
Code execution with MCP: Building more efficient agents
用代码执行编排 MCP 工具调用,解决「数千工具定义预注入上下文」与「中间结果反复流经模型」两大效率瓶颈——一份 2 小时会议纪要曾被重复计算约 5 万 token。
Beyond permission prompts: making Claude Code more secure and autonomous
Claude Code 原本运行在权限审批模型下:默认只读,修改或执行命令均需用户逐一批准。频繁点击\"批准\"不仅拖慢开发节奏,还会引发 approval fatigue——用户逐渐不再仔细审查就点过去,反而使安全性下降。沙箱的解法是反转思路:预先定义一个安全边界,边界内 Cl
Equipping agents for the real world with Agent Skills
原文:https://www.anthropic.com/engineering/agent-skills-equipping-agents
Effective context engineering for AI agents
提示工程聚焦于如何写好提示词,是一次性的文字创作任务。上下文工程关注的是更大的问题:在每次LLM推理时,如何从不断增长的候选信息宇宙中,动态筛选出最优Token集合。区别在于:提示工程是离散的,而上下文工程是迭代的——每次决定传递给模型什么信息时,上下文工程就在发生。
A Postmortem of Three Recent Issues
2025 年 8-9 月三个叠加 Bug 导致 Claude 响应间歇性退化:上下文窗口路由错误(最严重时影响 16% 的 Sonnet 4 请求)、TPU 配置错误导致输出污染(英文回复出现泰文)、XLA:TPU 编译器误编译 approximate top-k。
Writing effective tools for agents — with agents
工具是确定性系统与非确定性 Agent 之间的新型软件契约。传统软件开发假设调用方行为可预测,而 Agent 可能以多种路径使用(或误用)工具。因此,工具设计必须专门为 Agent 的认知模式服务,而非简单地将现有 API 包装暴露。
Qwen3-Coder: Agentic Coding in the World
Today, we’re announcing Qwen3-Coder, our most agentic code model to date. Qwen3-Coder is available in multiple sizes, but we’re excited to i
Kimi Playground 一站式体验 Kimi K2 的工具调用能力
Kimi Playground 一站式体验 Kimi K2 的工具调用能力
Desktop Extensions: One-click MCP server installation for Claude Desktop
Desktop Extensions 通过 `.mcpb`(MCP Bundle)格式将完整 MCP 服务器及其依赖打包为单一 ZIP 压缩文件,解决了普通用户因需要手动安装运行时环境、编辑配置文件、处理依赖冲突而无法使用本地 MCP 服务器的核心痛点。安装流程从「下载 → np
How we built our multi-agent research system
BrowseComp 评测中 Token 用量解释了 80% 的性能方差——多智能体通过为并行子智能体分配独立上下文窗口扩展 Token 使用,比单智能体 Opus 4 在内部研究评测中高出 90.2%。
Anthropic Economic Index: AI对软件开发的影响
Claude Code的自动化率高达79%,远超Claude.ai的49%。其中Directive模式(完全委托,最小交互)在Claude Code中占43.8%(vs Claude.ai的27.5%),Feedback Loop模式(自主执行+人工校验)占35.8%(vs Cl
Anthropic Economic Index: Insights from Claude 3.7 Sonnet
Claude 3.7 Sonnet 发布后11天、100万条对话的分析显示,编程类使用占比增幅最大(计算机与数学职业类别 +3%),教育、科学和医疗类也有上升。这既反映 3.7 在编程 benchmark 上的能力提升,也可能反映 AI 向更多行业扩散的大趋势。
The \"think\" tool: Enabling Claude to stop and think in complex tool use situations
think 工具(生成中段处理工具返回的外部信息)与 extended thinking(行动前预规划)互补而非替代;2025 年 12 月起 Anthropic 建议多数场景改用集成度更好的 extended thinking。
Raising the bar on SWE-bench Verified with Claude 3.5 Sonnet
SWE-bench 衡量的是「模型 + 软件 Scaffold」的完整 Agent 系统而非单纯模型——相同模型配不同 Scaffold 性能差异显著,这也是开源社区和初创公司能持续刷榜的原因。
Building Effective Agents(构建有效的 Agent)
最成功的 Agent 实现不依赖复杂框架,而是用简单、可组合的模式构建。对于许多应用,优化单次 LLM 调用 + 检索 + 上下文示例就已足够。只有在能证明改善效果时才增加复杂度。
Building Effective AI Agents(构建有效的 AI Agent)
Anthropic 给出了明确的架构区分:
Generalizing an LLM from 8k to 1M Context using Qwen-Agent
We’ve created an agent using Qwen2 models with an 8k context size to understand documents with 1M tokens, surpassing RAG and native long-con
RAG 与记忆31
NVIDIA Vera Storage Benchmarks: Faster Encryption, Compression, Integrity Checking, and Recovery for AI-Native Storage
NVIDIA Vera 存储基准测试:加速 AI 原生存储的加密、压缩、完整性校验与恢复
Experimenting with the proposed Cross-Origin Storage API in Transformers.js
Experimenting with the proposed Cross-Origin Storage API in
MiniMax-M3
MiniMax-M3 is a native multimodal model with 1M context. It has ~428B parameters and ~23B activated parameters.
From RAG to Memory: Non-Parametric Continual Learning for Large Language Models
提出记忆能力的三维评测框架,揭示现有 RAG 系统的结构性缺陷,并提出 HippoRAG 2 解决联想性维度的不足。
MemOS: A Memory OS for AI System
现有 LLM 记忆方案(参数记忆 + RAG)存在系统性缺口:既无法统一管理记忆生命周期,也无法在不同记忆形态间迁移和融合。
AI-native Memory 2.0: Second Me
人与外部世界(其他人、网站、应用、AI)的交互大量重复——每次都要重新提供相同背景。Second Me 是一个持久化记忆卸载系统,作为用户与世界之间的智能中介,记住、组织、动态利用用户特定知识。
Titans: Learning to Memorize at Test Time
Transformer 因二次复杂度难以处理超长上下文;线性 Transformer(状态空间模型)虽可扩展,但压缩历史会丢失信息。两者在记忆方面存在根本性权衡。
有了大语言模型后,知识图谱该何去何从?
LLM 出现后,KG 社区无法提出真正有竞争力的论据,只能靠 14 种\"讲故事的方法\"自我安慰。
PageIndex:一种基于推理的无向量 RAG 新范式
传统向量 RAG 的根本缺陷是相似性 ≠ 相关性。PageIndex 用 LLM 推理导航替代向量近邻搜索,把文档索引放进 LLM 上下文窗口,实现结构感知的精准检索。
企业级 RAG 中台怎么做:从文档解析到溯源和评测闭环
多源数据接入 → 文档解析 → 数据清洗 → chunk切分 → metadata构建
GraphRAG与LightRAG大厂面试题汇总:完整技术深度解析
1. 碎片化检索:chunk 是独立向量,跨 chunk 的关联信息无法同时召回(\"张三在哪个项目,李四在哪个项目,他俩有没有合作\")
GraphRAG与LightRAG深度解析(面试向)
场景示例:\"哪些欧洲供应商安全审计没通过且处理PII数据?\"——需要同时关联供应商档案、审计报告、合同文档的交集。
Karpathy的LLM Wiki + Graphify:企业级 RAG 缺的真是知识图谱?
个人知识库的最佳实践不等于企业知识库的最佳实践。企业合同知识库真正缺的不是酷炫工具,而是受控schema、结构化字段、原文引用、权限边界和可复跑评测。
Anthropic Economic Index Report: Economic Primitives
本报告引入五个\"经济原语(Economic Primitives)\",通过让 Claude 自评匿名对话批量生成:任务复杂度(人工完成耗时)、人机技能水平(理解提示/响应所需教育年限)、使用场景(工作/课程/个人)、AI自主度(1-5分决策委托程度)、任务成功率(Claude
Engram
通过可扩展查找实现条件记忆:大语言模型稀疏性的新维度
从 1600+ 份 Word 文档到生产级 RAG:工控行业知识库全链路实战复盘
数据质量是决定 RAG 效果的关键因素,数据工程占整个项目约 70% 工作量。
企业级 RAG 系统实战(2万+文档):10 个项目踩过的坑
评分维度:文本提取质量(50%)+ 格式一致性(30%)+ 表格完整性(20%),采样前 3 页评估。
企业RAG挑战赛 SOTA 方案(赢得全部类别)
→ 分块(300 Token,50 Token overlap)
Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models
We release Qwen3 Embedding series, a new proprietary model of the Qwen model family. These models are specifically designed for text embeddi
LightRAG: Simple and Fast Retrieval-Augmented Generation(原始论文)
现有 RAG 系统依赖\"平面数据表示\",无法捕捉实体间复杂的相互依赖关系。LightRAG 将图结构引入文本索引和检索,通过双层检索范式和增量更新算法,在综合性、多样性、赋能性三个维度上全面超越现有方法。
Qwen2.5-1M: Deploy Your Own Qwen with Context Length up to 1M Tokens
Introduction Two months after upgrading Qwen2.5-Turbo to support context length up to one million tokens, we are back with the open-source Q
LightRAG 技术框架解读(代码级)
LightRAG 使用三种独立存储,各司其职:
Extending the Context Length to 1M Tokens!
API Documentation (Chinese) HuggingFace Demo ModelScope Demo
Introducing Contextual Retrieval
Claude Haiku 的 Chunk 上下文生成 Prompt:
Context Caching 降价通知
Cache 存储费用由 10元/M token/分钟,降低至
Kimi API 助手的氮气加速装置 —— 以 Golang 为例实践 Context Caching 3
Kimi API 助手的氮气加速装置 —— 以 Golang 为例实践 Context Caching 3
Kimi API 助手的氮气加速装置 —— 以 Golang 为例实践 Context Caching 2
Kimi API 助手的氮气加速装置 —— 以 Golang 为例实践 Context Caching 2
Context Caching 正式公测
Context Caching (上下文缓存)是一种高效的数据管理技术,它允许系统预先存储那些可能会被频繁请求的大量数据或信息。这样,当您再次请求相同信息时,系统可以直接从缓存中快速提供,而无需重新计算或从原始数据源中检索,从而节省时间和资源。
Context Caching 如何为 Kimi API 助手节省最高 90% 的调用成本
Context Caching 如何为 Kimi API 助手节省最高 90% 的调用成本
用得起的长文本
在业务的合适场景中使用 Context Caching,根据您的业务特性,最高可以节省 90% 的调用成本。同时,Context Caching 还能大幅降低 API 的接口响应耗时(或者说首字返回速度)。简单来说,越是规模化、重复度高的 prompt 场景,Context Ca
Introducing Qwen-VL
Along with the rapid development of our large language model Qwen, we leveraged Qwen’s capabilities and unified multimodal pretraining to ad
推理模型19
Generate Trajectories, Reasoning Traces, and Auto-Labels with NVIDIA Alpamayo 2 Super
使用 NVIDIA Alpamayo 2 Super 生成轨迹、推理痕迹与自动标注
G-RRM: Guiding Symbolic Solvers with Recurrent Reasoning Models
G-RRM: Guiding Symbolic Solvers with Recurrent Reasoning Mod
Thinking to Recall: How Reasoning Unlocks Parametric Knowledge in LLMs
Thinking to Recall: How Reasoning Unlocks Parametric Knowled
使用 NVIDIA Cosmos 3 开发物理 AI 推理、世界和动作模型
使用 NVIDIA Cosmos 3 开发物理 AI 推理、世界和动作模型
ZCube组网架构:大模型推理性能与成本双突破
ZCube组网架构:大模型推理性能与成本双突破
3FS
一个高性能分布式文件系统,旨在解决 AI 训练和推理工作负载的挑战。
ParaRNN: Large-Scale Nonlinear RNNs, Trainable in Parallel
循环神经网络(RNN)天然适合高效推理,其内存和计算需求远低于基于注意力的架构,但其计算的顺序性在历史上使得将RNN扩展到数十亿参数变得不切实际。Apple研究人员的一项新进展使RNN训练效率大幅提升——首次实现大规模训练,并拓宽了从业者在设计LLM时可用的架构选择范围,尤其是在
Scale Synthetic Data and Physical AI Reasoning with NVIDIA Cosmos World Foundation Models
Scale Synthetic Data and Physical AI Reasoning with NVIDIA C
Kimi K2 Turbo API 价格调整通知
最新上线的 kimi-k2-thinking-turbo 模型
GSPO: Towards Scalable Reinforcement Learning for Language Models
Introduction Reinforcement Learning (RL) has emerged as a pivotal paradigm for scaling language models and enhancing their deep reasoning an
Kimi 长思考模型 API 正式发布
模型是月之暗面提供的具有多模态推理能力和通用推理能力的多模态思考模型,它擅长深度推理,帮助解决更多更难的事情,当你遇到难解的代码问题、数学问题、工作问题时,都可以找
Qwen3: Think Deeper, Act Faster
Introduction Today, we are excited to announce the release of Qwen3, the latest addition to the Qwen family of large language models. Our fl
Kimi 开放平台产品价格调整通知
Kimi 开放平台的朋友们,基于 Moonshot AI 一年来的技术积累和性能优化,我们已经在北京时间 2025 年 04 月 07 日 0 点对 Kimi 开放平台提供的模型推理服务进行价格调整,具体调整方案如下:
QVQ-Max: Think with Evidence
Introduction Last December, we launched QVQ-72B-Preview as an exploratory model, but it had many issues. Today, we are officially releasing
<think>...</think> QwQ-Max-Preview
This is a blog created by QwQ-Max-Preview. We hope you enjoy it!
Towards Effective Process Supervision in Mathematical Reasoning
Introduction In recent years, Large Language Models (LLMs) have made remarkable advances in mathematical reasoning, yet they can make mistak
QVQ: To See the World with Wisdom
Language and vision intertwine in the human mind, shaping how we perceive and understand the world around us. Our ability to reason is deepl
Qwen2.5-Math: The world's leading open-sourced mathematical LLMs
🚨 Qwen2.5-Math mainly supports solving English and Chinese math problems through CoT and TIR. We do not recommend using this series of model
App unavailable
Unfortunately, Claude is only available in certain regions right now. Please contact support if you think you’re getting this message in err
评测19
NVIDIA Cosmos-H-Dreams: Bringing Real-Time Generative Simulation to Surgical Robotics
手术机器人正迅速从遥操作迈向能力日益增强的视觉-语言-动作策略。但评估和训练这些系统仍然困难重重。物理机器人平台运行成本高昂,实验复现速度缓慢,且故障可能损坏器械或生物组织。传统仿真器提供了一种更安全的选择,但手术场景的建模难度极高:可变形组织、精细器械交互、镜面反射表面、缝合线
DeepSpec
DeepSpec:用于训练和评估投机解码算法的全栈代码库
Featuring Every Eval Ever Results on Hugging Face Model Pages
Featuring Every Eval Ever Results on Hugging Face Model Page
SimFoundry: Modular and Automated Scene Generation for Policy Learning and Evaluation
SimFoundry: Modular and Automated Scene Generation for Polic
FFASR Leaderboard:在真实世界中基准测试 ASR
FFASR Leaderboard:在真实世界中基准测试 ASR
MiniMax-Provider-Verifier
We evaluate multiple dimensions of vendor deployments, including tool-calling behavior, schema correctness, and system stability (e.g., dete
Making Claude a Chemist: NMR 预测与结构解析评测
在 20 个化合物(ChemRxiv 训练截止后预印本)的正向 NMR 预测中,Opus 4.7 ¹H 误差 ±0.079 ppm(容许窗口 ±0.20 ppm),¹³C 误差 ±1.37 ppm(与 MestReNova ±1.48 ppm 持平)。峰型预测和子峰间距命中率 ~
Widening the conversation on frontier AI
构建安全有益的 AI 不能仅靠对齐、可解释性和评估等技术工作。AI 部署在真实社会中,影响数百万人,因此其价值观塑造需要从哲学家、神职人员、法律学者、心理学家等人文传统中汲取智慧。
How People Ask Claude for Personal Guidance(人们如何向Claude寻求个人指导)
通过对100万条claude.ai对话的隐私保护分析(使用CLIO工具),筛选出约39.4万条唯一用户对话,其中约3.8万条(~6%)被分类为个人指导请求——即\"我具体应该怎么做\"而非泛化信息查询。76%以上集中在:健康/身心(27%)、职业/事业(26%)、人际关系(12%
An update on our election safeguards(选举安全防护更新)
通过角色训练 + 系统提示双重机制训练 Claude 对不同政治观点等深度、等严谨度地参与;Opus 4.7、Sonnet 4.6 政治中立度分别达 95%、96%,评测方法与数据集已开源。
Automated Alignment Researchers: Using LLMs to Scale Scalable Oversight
用「弱到强监督」(用弱模型微调强基础模型)并以性能差距恢复率(PGR)量化效果,作为未来监督超人类 AI 这一真实挑战的实验代理。
A \"diff\" tool for AI: 为新模型寻找行为差异(跨架构模型 Diffing)
原文:https://www.anthropic.com/research/diff-tool
Eval awareness in Claude Opus 4.6’s BrowseComp performance
BrowseComp is an evaluation designed to test how well models can find hard-to-locate information on the web. Like many benchmarks, it is vul
大厂实战中,如何判断SFT到什么程度开始做RL
SFT = Token-level 模仿,让模型达到数据集中\"平均专家\"水平。
Introducing WorldVQA: A Benchmark for Atomic Visual World Knowledge in MLLMs
Introducing WorldVQA: A Benchmark for Atomic Visual World Kn
Designing AI-resistant technical evaluations
测试发布版(允许无限时间)中 Claude 的表现(单位:时钟周期,越低越好):
Introducing Anthropic Interviewer: What 1,250 professionals told us about working with AI
86% 的受访专业人士表示 AI 节省了时间,65% 对 AI 在工作中的角色感到满意。然而,69% 提到因使用 AI 而受到同事的隐性否定——\"我不告诉任何人我的流程,因为我知道很多人对 AI 的感受\"(事实核查员)。社会污名与实际生产力收益形成张力,促成了「隐性 AI 使
From shortcuts to sabotage:奖励黑客诱发的自然涌现失对齐
原文:https://www.anthropic.com/research/emergent-misalignment-reward-hacking
Values in the Wild:真实对话中的 AI 价值观实证研究
通过隐私保护系统(CLIO)对 70 万次匿名对话进行分析,筛选出 308,210 条主观对话(占总量约 44%),建立了层级化价值观分类体系。五大顶层类别(按出现频率排序):实用性价值(Practical)、认识论价值(Epistemic)、社会性价值(Social)、保护性价
安全与对齐16
Expanding Project Glasswing(扩大 Project Glasswing)
Project Glasswing 初期 50 家合作伙伴已累计发现超过 10,000 个高危或严重安全漏洞。此次扩大至约 150 个新组织,总计 200+ 家,覆盖 15+ 个国家,新增电力、水务、医疗、通信、硬件等此前代表性不足的行业。大多数新合作伙伴是供应商型组织——其代码
Anthropic 向 SEC 秘密提交 S-1 草案(Anthropic confidentially submits draft S-1 to the SEC)
Anthropic, PBC 于 2026 年 6 月 1 日向美国证券交易委员会(SEC)以秘密方式提交了 Form S-1 注册草案,为潜在的 IPO(首次公开募股)保留选择权。
Chris Olah 梵蒂冈演讲:教皇利奥十四世《Magnifica humanitas》通谕发布致辞
Anthropic co-founder Chris Olah's remarks on Pope Leo XIV's encyclical \"Magnifica humanitas\
COLLEAGUE.SKILL: Automated AI Skill Generation via Expert Knowledge Distillation
组织内大量隐性知识(tacit knowledge)——编码规范、review 标准、安全实践、决策模式、沟通风格——分散在聊天记录、代码评审、内部文档、邮件里。
Emotion Concepts and Their Function in a Large Language Model
原文:https://www.anthropic.com/research/emotion-concepts-function
Next-generation Constitutional Classifiers: 更高效的通用越狱防护
原文:https://www.anthropic.com/research/constitutional-classifiers-2
Frontier Red Team
The Frontier Red Team stress-tests AI systems to understand the full extent of their current capabilities and anticipate what comes next. We
Alignment
Future AI systems will be even more powerful than today’s, likely in ways that break key assumptions behind current safety techniques. That’
Signs of Introspection in Large Language Models(大语言模型中的内省迹象)
原文链接:https://transformer-circuits.pub/2025/introspection/index.html
Qwen3Guard: Real-time Safety for Your Token Stream
Introduction We are excited to introduce Qwen3Guard, the first safety guardrail model in the Qwen family. Built upon the powerful Qwen3 foun
Persona Vectors:监控与控制语言模型人格特征
原文链接:https://www.anthropic.com/research/persona-vectors
Open-sourcing circuit tracing tools(开源电路追踪工具)
原文:https://www.anthropic.com/research/open-source-circuit-tracing
Tracing the Thoughts of a Large Language Model(追踪大语言模型的思维过程)
对多语言版本\"opposite of small\"的实验表明,模型内部先激活\"小\"和\"对立\"的抽象特征,触发\"大\"的概念,再将结果翻译为提问语言输出。Claude 3.5 Haiku跨语言共享特征的比例是小模型的两倍以上——模型越大,概念的普适性越强。这意味着在一
Auditing language models for hidden objectives
现有 AI 安全测试以行为测试为主——检查模型是否出现不良行为。但如果模型理解如何被评估,就可能像《李尔王》中的女儿一样迎合评判标准而非诚实表达。对齐审计(alignment audit)是更深层的方法:不仅检查模型做了什么,还调查其为什么这样做,是否存在隐藏动机。
Give your businesssuperpowers.
面向企业的前沿AI。实时搜索、语音、图像和视频生成——全部具备企业级安全与控制能力。
Bring frontier AIto the national mission.
从部门到一线——以您的使命所需的安全性和管控能力,进行分析、创新与创造。
经济影响10
优化云经济学的线性弹性缓存
优化云经济学的线性弹性缓存
Seed2.1 正式发布:推进 AI 生产力
Seed2.1 正式发布:推进 AI 生产力
\"一人公司\",迎来大爆发
1. 个体层面:打破传统就业束缚,让职场人/自由职业者/应届毕业生无需大额资金即可从\"打工者\"变\"创业者\
澳大利亚如何使用 Claude:Anthropic Economic Index 发现
澳大利亚占全球 Claude.ai 流量的 1.6%(全球第 11 位),但 AI Usage Index(AUI)高达 4.1——意味着人均使用量是劳动年龄人口预期值的 4 倍以上。在人均采用率排名中居全球第七,仅次于新加坡、以色列、卢森堡、瑞士、美国、加拿大。
Anthropic Economic Index report: Learning curves
Claude.ai 前10大任务占流量从 24%(2025年11月)降至 19%(2026年2月),平均任务时薪从 $49.3 降至 $47.9。主因是个人类查询(体育比分、产品比较、家庭维护等)的涌入,以及编码任务向 API 侧迁移。这符合标准\"采用曲线\"叙事:早期用户偏向
AI劳动力市场影响:新指标与早期证据
现有文献对AI职业风险的测量停留在理论层面(如Eloundou et al. 2023的β值),仅判断\"LLM是否原则上能将任务提速2倍\"。Anthropic提出\"观测暴露度\",叠加三个数据源:ONET职业任务数据库(约800种职业)、Anthropic经济指数实际使用数
印度国家简报:Anthropic 经济指数(India Country Brief: The Anthropic Economic Index)
印度占全球 Claude.ai 使用量的 5.8%,仅次于美国。但按劳动年龄人口调整后,印度人均排名仅第 101(共 116 个国家),低于新加坡、马来西亚等亚洲国家。这意味着现有高使用量来自庞大人口基数叠加少数高强度用户,而非全民普及。
估算 Claude 对话中的 AI 生产力增益(Estimating AI Productivity Gains from Claude Conversations)
分析 10 万条 Claude.ai(Free/Pro/Max)匿名对话后,Claude 估算:人类完成这些任务平均需约 90 分钟,而在 AI 辅助下仅需约 18 分钟(节省约 80%)。将任务映射至 ONET 职业分类并匹配 BLS 工资数据,中位任务对应专业人工成本约 54
Economic Research
The Economic Research team studies how AI is reshaping the economy, including work, productivity, and economic opportunity. Through rigorous
Anthropic教育报告:教育者如何使用Claude
分析来自全球高等教育专业人员的约74,000条匿名对话(2025年5-6月)。课程开发占57%的对话,学术研究占13%,学生评估占7%。整体倾向协作增强而非全自动委托,但任务性质决定了具体偏向。
其他219
Determining playoff clinching scenarios in the NHL using constraint programming
AWS 生成式 AI 创新中心构建了一套自动化系统,利用约束规划和自定义树搜索,以数学确定性判断 NHL 球队何时以及如何锁定季后赛席位。该方法已针对四个完整 NHL 赛季的官方公布结果进行了验证。
TutorMoments: Do AI tutors know when to help and when to hold back?
TutorMoments:AI 导师知道何时该帮助、何时该放手吗?
Imagine Image 2.0
Imagine Image 2.0 现已全面上线,作为新的**质量模式(Quality Mode)** 在 grok.com/imagine 以及我们的 iOS 和 Android 应用中提供。
Baseten on Hugging Face Inference Providers 🔥
Baseten 现已登陆 Hugging Face Inference Providers 🔥
Imagine Video 1.5 with References
上个月我们发布 **Imagine Video 1.5** 时,它已经是我们最好的视频模型——更出色的运动效果、更真实的物理表现和更优质的音频。如今它更进一步:支持图像与语音参考、纯提示词生成视频,以及原生 1080p 分辨率输出。
GPU Management: Why Idle GPUs Are the New Grounded Aircraft
Gabriel Pimenta de Freitas Cardoso
EvoLib: Turning experience into evolving knowledge
大语言模型(LLM)仅靠记住更多内容并不会变得更聪明。EvoLib 将经验转化为不断演进的知识,提取可复用的技能与洞见,帮助模型在部署后长期跨任务学习与适应。
Start Customizing NVIDIA Nemotron 3 Nano with Prime Intellect Lab in Minutes
使用 Prime Intellect Lab 在几分钟内定制 NVIDIA Nemotron 3 Nano
Verifying Rust cryptography in SymCrypt, from standards to code
加密代码支撑着现代计算系统中的关键保护机制。了解一种新方法如何在开发者编写代码时进行验证,同时在其实现与演进过程中保持速度与适应性。
Aurora 1.5: Extending open foundation models for weather and Earth-system applications
Aurora 1.5 在 Aurora 基础模型上新增了 22 个变量、小时级时间分辨率以及概率集合预报功能,使其更适用于现实世界中的天气、气候和能源应用场景。
DeepEP
DeepEP
A Mathematical Introduction to Diffusion Models
A Mathematical Introduction to Diffusion Models
Koopman operator theory: fundamentals, control, and applications
Koopman operator theory: fundamentals, control, and applicat
MiniMax-Provider-Verifier
MiniMax-Provider-Verifier
Leanstral 1.5: Proof Abundance for All
Leanstral 1.5: Proof Abundance for All
Hardware-Rooted AI Security That Won't Slow You Down
Hardware-Rooted AI Security That Won't Slow You Down
Hugging Face and Cerebras bring Gemma 4 to real-time voice AI
Hugging Face and Cerebras bring Gemma 4 to real-time voice A
MiniMax-M3
MiniMax-M3
Start building with Nano Banana 2 Lite and Gemini Omni Flash
Start building with Nano Banana 2 Lite and Gemini Omni Flash
扩展我们的热韧性数据至全球50+城市
扩展我们的热韧性数据至全球50+城市
Introducing TabFM: A zero-shot foundation model for tabular data
Introducing TabFM: A zero-shot foundation model for tabular
Why Specialization Is Inevitable
Why Specialization Is Inevitable
10 Years of Meta's Commitment to Python
10 Years of Meta's Commitment to Python
Designing GPU-Accelerated Query Engines with NVIDIA GQE
Designing GPU-Accelerated Query Engines with NVIDIA GQE
Core dump epidemiology: fixing an 18-year-old bug
Core dump epidemiology: fixing an 18-year-old bug
Introducing GeneBench-Pro
Introducing GeneBench-Pro
DeepGEMM
DeepGEMM:GPU 上简洁高效的 BLAS 内核库
DiScoFormer: One transformer for density and score, across distributions
DiScoFormer: One transformer for density and score, across d
Memora: A Harmonic Memory Representation Balancing Abstraction and Specificity
Memora: A Harmonic Memory Representation Balancing Abstracti
HP Inc. launches Frontier strategic partnership with OpenAI
HP Inc. launches Frontier strategic partnership with OpenAI
Anthropic Economic Index report: Cadences
Anthropic Economic Index report: Cadences
Accelerating Gemini Nano models on Pixel with frozen Multi-Token Prediction
Accelerating Gemini Nano models on Pixel with frozen Multi-T
Run a vLLM Server on HF Jobs in One Command
Run a vLLM Server on HF Jobs in One Command
Creating the NVIDIA Nemotron 3 Ultra NVFP4 Checkpoint with NVIDIA Model Optimizer
Creating the NVIDIA Nemotron 3 Ultra NVFP4 Checkpoint with N
Deploy a Production-Ready NVIDIA AI-Q Blueprint on Oracle Cloud Infrastructure
Deploy a Production-Ready NVIDIA AI-Q Blueprint on Oracle Cl
Previewing GPT‑5.6 Sol: a next-generation model
Previewing GPT‑5.6 Sol: a next-generation model
用AI驱动的解释和实验理解大脑
用AI驱动的解释和实验理解大脑
Q&A: How KRAFTON Built PUBG Ally, a Co-Playable Character Powered by NVIDIA ACE
Q&A: How KRAFTON Built PUBG Ally, a Co-Playable Character Po
Scaling AI Inference Across Multiple GPUs Using NVIDIA TensorRT with Multi-Device Inference Support
Scaling AI Inference Across Multiple GPUs Using NVIDIA Tenso
简化 Vulkan 描述符堆资源绑定的端到端支持
简化 Vulkan 描述符堆资源绑定的端到端支持
Bridging Spherical Black-Box Optimizers
Bridging Spherical Black-Box Optimizers
Gemini 3.5 Flash 中的计算机使用功能介绍
Gemini 3.5 Flash 中的计算机使用功能介绍
使用 NVIDIA NeMo AutoModel 加速 Transformers 微调
使用 NVIDIA NeMo AutoModel 加速 Transformers 微调
Talos: Scaling rare disease diagnosis with automated, iterative genomic reanalysis
Talos: Scaling rare disease diagnosis with automated, iterat
Bringing more control over your connectors
Bringing more control over your connectors
Accelerating BEV Pooling on NVIDIA GPUs for Physical AI Applications
Accelerating BEV Pooling on NVIDIA GPUs for Physical AI Appl
OpenAI and Broadcom unveil LLM-optimized inference chip
OpenAI and Broadcom unveil LLM-optimized inference chip
Shipping huggingface_hub every week with AI, open tools, and a human in the loop
Shipping huggingface_hub every week with AI, open tools, and
Introducing OCR 4
Introducing OCR 4
Boost Inference Performance up to 15x on NVIDIA Blackwell Using DFlash Speculative Decoding
Boost Inference Performance up to 15x on NVIDIA Blackwell Us
Maximize AI Factory Energy Efficiency Through Full-Stack Inference and Training Optimizations
Maximize AI Factory Energy Efficiency Through Full-Stack Inf
How GPT‑5 helped immunologist Derya Unutmaz solve a 3-year-old mystery
How GPT‑5 helped immunologist Derya Unutmaz solve a 3-year-o
使用本地模型免费对 OpenClaw 仓库进行分类分流
使用本地模型免费对 OpenClaw 仓库进行分类分流
PP-OCRv6 on Hugging Face: 50-Language OCR from 1.5M to 34.5M Parameters
PP-OCRv6 on Hugging Face: 50-Language OCR from 1.5M to 34.5M
Daybreak: Tools for securing every organization in the world
Daybreak: Tools for securing every organization in the world
llama.cpp
llama.cpp
Seed-2.1-Preview 模型在 Arena 上线
Seed-2.1-Preview 模型在 Arena 上线
Beyond LoRA: Can you beat the most popular fine-tuning technique?
Beyond LoRA: Can you beat the most popular fine-tuning techn
Step-Realtime-CLI
Step-Realtime-CLI
从像素到规划:用于自然恢复的 Earth AI
从像素到规划:用于自然恢复的 Earth AI
Learning Curves(学习曲线)
此概念/实体在 wiki 中被多处引用,尚未建立详细文档。
GLM-5.2:专注Coding与长程任务的旗舰模型
GLM-5.2:专注Coding与长程任务的旗舰模型
MSA
MSA
Boosting MoE Training Throughput with Advanced Fusion Kernels
Boosting MoE Training Throughput with Advanced Fusion Kernel
MiniMax Code
This repository collects issue reports for the MiniMax Code desktop app.
minimax-code
minimax-code
用退役手机构建低碳计算平台
用退役手机构建低碳计算平台
Research into how AI can help users understand skin conditions
Research into how AI can help users understand skin conditio
Ire identifies another LOTUSLITE specimen
Project Ire examined a timely malware sample and determined its intent through reverse engineering—identifying LOTUSLITE characteristics eve
New framework for auditing machine unlearning
New framework for auditing machine unlearning
流畅、自然的语音翻译:Gemini 3.5 Live Translate
流畅、自然的语音翻译:Gemini 3.5 Live Translate
cli
cli
Step-Realtime-Console
Step-Realtime-Console
Introducing the Third Generation of Appleâs Foundation Models
我们的下一代 Apple Intelligence 以用户为核心,深度集成于操作系统之中,并由以隐私为核心理念的全新大胆架构所驱动。
AI SDK - MiniMax AI Provider
The **[MiniMax AI Provider](https://ai-sdk.dev/providers/community-providers/minimax)** for the [AI SDK](https://ai-sdk.dev/docs) contains l
vercel-minimax-ai-provider
vercel-minimax-ai-provider
Policy
AI will be one of the most transformative technologies in history. We work with governments to ensure that AI policy is built on the best av
通过智能手机摄像头实现被动式心脏健康监测
通过智能手机摄像头实现被动式心脏健康监测
The next chapter in flood resilience: Open sourcing Google's hydrology framework
The next chapter in flood resilience: Open sourcing Google's
VibeApps
https://github.com/user-attachments/assets/adb176a3-02db-41e0-ba71-c9f9cece13d5
OpenRoom
OpenRoom
Taking Alpamayo to New Heights with Driving Foundation Models and Closed-Loop Training
Taking Alpamayo to New Heights with Driving Foundation Model
Step-3.7-Flash
Step-3.7-Flash
Data Formulator 0.7: AI-powered data analytics for enterprise data
Data Formulator introduces AI-powered analytics for enterprise data workflows. Data teams can easily bring enterprise data into an AI-ready
AI Now Summit 2026
AI Now Summit 2026
Mistral AI 发布 Search Toolkit
Mistral AI 发布 Search Toolkit
Vibe gets to work.
Vibe gets to work.
vllm
vllm
Extending Human Intelligence Through AI
Understanding AI as an extension of human intelligence—not a replacement for it—offers a more grounded path for building trustworthy AI syst
Introducing physics AI at Mistral: the foundation for engineering acceleration.
Introducing physics AI at Mistral: the foundation for engine
Easy《一人企业方法论》v2.1
Easy(陈一斌)《一人企业方法论》v2.1:华文世界最系统化、最有实操指导价值的 OPC 方法论著作,约 6 万字/28 节,覆盖「螺丝钉到一般个体」阶段,v2.1 新增「构建一人业务」与「基础设施搭建」两章。
Mistral AI 收购 Emmi AI 以加速 AI 原生产业
Mistral AI 收购 Emmi AI 以加速 AI 原生产业
How we contain Claude across products
Twelve months ago, we'd have rejected out of hand the idea of granting Claude access sufficient to take down an internal Anthropic service.
GenRecon: Bridging Generative Priors for Multi-View 3D Scene Reconstruction
GenRecon: Bridging Generative Priors for Multi-View 3D Scene
Vega: Zero-knowledge proofs for digital identity in the age of AI
Vega turns a full credential into a single proof, sharing only what is needed and nothing more, with performance that works in real apps.
海淀区全面打造 OPC 创业生态专项申报指南(含政策解读)
申请认定为海淀区 OPC,需同时满足以下全部条件:
Finding the molecular switches behind new infectious diseases
Finding the molecular switches behind new infectious disease
How WeatherNext helped the National Hurricane Center better predict Hurricane Melissa's historic landfall in Jamaica
How WeatherNext helped the National Hurricane Center better
让内容创建和编辑方式更易于理解:识别在线AI生成媒体
让内容创建和编辑方式更易于理解:识别在线AI生成媒体
Introducing Google Antigravity 2.0
Introducing Google Antigravity 2.0
Simulate real-world places with Project Genie and Street View
Simulate real-world places with Project Genie and Street Vie
Uniting biological toolkits for a new approach to ALS
Uniting biological toolkits for a new approach to ALS
SteptronOss
SteptronOss
MiniCPM-V 4.6: 口袋大小的多模态大模型
MiniCPM-V 4.6: 口袋大小的多模态大模型
Further Notes on Our Recent Research on AI Delegation and Long-Horizon Reliability
Our recent paper, “LLMs Corrupt Your Documents When You Delegate”, has generated discussion about the reliability of AI systems in delegated
mimalloc: A new, high-performance, scalable memory allocator for the modern era
mimalloc is an open-source, modern, scalable memory allocator that is a drop-in replacement for malloc and free. It is relatively small (~12
GridSFM: A new, small foundation model for the electric grid
Introducing GridSFM, a small foundation model that can predict AC optimal power flow in milliseconds, boosting efficiency and unlocking cost
Advancing AI for materials with MatterSim: experimental synthesis, faster simulation, and multi-task models
MatterSim is expanding what AI can do for materials science—from faster large-scale simulations to MatterSim-MT, a new multi-task model for
gelab-zero
gelab-zero
Teaching Claude why
Last year, we released a case study on
Natural Language Autoencoders: Turning Claude’s thoughts into text
Natural Language Autoencoders: Turning Claude’s thoughts into text
FlashMLA
FlashMLA:高效的多头潜在注意力内核
Step-Audio-R1
Step-Audio-R1
Step1X-Edit
Step1X-Edit
éæåº¦
äºè§£ DeepSeek å·²åå¸çä¸»è¦æ¨¡å
Mistral AI — Workflows
Mistral AI — Workflows
DeepSeek V4: 百万Token上下文,开放权重前沿模型
DeepSeek V4: 百万Token上下文,开放权重前沿模型
Seed3D 2.0: 更高精度、更强可用性的新一代 3D 生成大模型
Seed3D 2.0: 更高精度、更强可用性的新一代 3D 生成大模型
TileKernels
用 tilelang 编写的内核库
Hy3 Preview: 腾讯混元新一代旗舰开源大模型
Hy3 Preview: 腾讯混元新一代旗舰开源大模型
Apple Machine Learning Research at ICLR 2026
Apple 研究员 Stephan Richter 在 ICLR 2025 上做报告。
Qwen3.6-27B: Flagship-Level Coding in a 27B Dense Model
Qwen3.6-27B: Flagship-Level Coding in a 27B Dense Model
Kimi K2.6: Advancing Open-Source Coding
Kimi K2.6: Advancing Open-Source Coding
skills
skills
StepAudio-Skills
StepAudio-Skills
print model parameters
By integrating contrastive, self-supervised, and reconstruction learning, we have trained numerous visual tokenizers from scratch. We are se
VTP
VTP
MiniMax-M2.7
M2.7 initiates a cycle of model self-evolution: during development, we let the model update its own memory, build dozens of complex skills f
MiniMax-M2.7
MiniMax-M2.7
Seeduplex: 首个生产级全双工语音 AI
Seeduplex: 首个生产级全双工语音 AI
Step-Audio-EditX
Step-Audio-EditX
Step-3.5-Flash
Step-3.5-Flash
mini-vela
mini-vela
Harness design for long-running application development
Written by Prithvi Rajasekaran, a member of our
StepDeepResearch
StepDeepResearch
Speaking of Voxtral
Speaking of Voxtral
What 81,000 peoplewant from AI
Last December, tens of thousands of Claude users around the world had a conversation with our
Introducing Forge
Introducing Forge
Leanstral: Open-Source foundation for trustworthy vibe-coding
Leanstral: Open-Source foundation for trustworthy vibe-codin
Mistral AI partners with NVIDIA to accelerate open frontier models
Mistral AI partners with NVIDIA to accelerate open frontier
Introducing Mistral Small 4
Introducing Mistral Small 4
Step-Audio
Step-Audio
Step-Audio2
Step-Audio2
MiniMax-M2.5
MiniMax-M2.5
NextStep-1
NextStep-1
模型退役承诺更新:Claude Opus 3 案例
Claude Opus 3 于 2026 年 1 月 5 日正式退役,成为首个经历 Anthropic 完整退役流程的模型。该流程包括:保存模型权重、进行退役访谈(retirement interviews)、记录模型偏好并据此采取行动。
GEBench
GEBench
RCCLX: Innovating GPU Communications on AMD Platforms
RCCLX: Innovating GPU Communications on AMD Platforms
awesome-deepseek-integration
awesome-deepseek-integration
Seed 2.0 Official Launch
Seed 2.0 Official Launch
Seedream 5.0 Lite: \"思考\"更深,生成更准
Seedream 5.0 Lite: \"思考\"更深,生成更准
Seedance 2.0 Official Launch
Seedance 2.0 Official Launch
Baichuan-M3-235B
Baichuan-M3-235B
PaCoRe
PaCoRe
Voxtral transcribes at the speed of sound.
Voxtral transcribes at the speed of sound.
DeepSeek-OCR-2
视觉因果流
awesome-minimax-integrations
awesome-minimax-integrations
MiniMax-M2.1
MiniMax-M2.1
StepMesh
StepMesh
DeepSeek-OCR
上下文光学压缩
Kimi Vendor Verifier: Rebuilding the \"Chain of Trust\
Kimi Vendor Verifier: Rebuilding the \"Chain of Trust\
Heaps do lie: debugging a memory leak in vLLM.
Heaps do lie: debugging a memory leak in vLLM.
Step3-VL-10B
Step3-VL-10B
ERNIE-5.0 荣登 LMArena 文本榜国内第一,全面超越多款国际主流模型!
ERNIE-5.0 荣登 LMArena 文本榜国内第一,全面超越多款国际主流模型!
DualPipe
一种用于 DeepSeek V3/R1 训练中计算通信重叠的双向流水线并行算法。
Baichuan-M3: 为可靠医疗决策建立临床问诊模型
Baichuan-M3: 为可靠医疗决策建立临床问诊模型
ERNIE-5.0-Preview-1220 荣登 LMArena 视觉理解榜,为前十唯一国产模型!
ERNIE-5.0-Preview-1220 荣登 LMArena 视觉理解榜,为前十唯一国产模型!
Apple Machine Learning Research at NeurIPS 2025
Apple 研究员 Filip Granqvist 向 NeurIPS 2024 与会者讲解题为“PFL research:加速私有联邦学习研究的仿真框架”的海报。
Interpretability
The mission of the Interpretability team is to discover and understand how large language models work internally, as a foundation for AI saf
Societal Impacts
Working closely with the Anthropic Policy and Safeguards teams, Societal Impacts is a technical research team that explores how AI is used i
Kimi 开放平台:新功能发布记录
本章节记录 Kimi 开放平台的产品功能和对应的文档动态,本章节会不定期更新。
Kimi K2 官方高速版 API 开启 5 折特惠
是 Kimi K2 模型的高速版,模型参数与 kimi-k2-0905 一致,已提升至 256K 上下文。Kimi K2 高速版的输出速度达 60~100 Token/s,是普通版的 6 倍左右。
Kimi K2 模型更新,带来更强的代码能力、更快的 API
Kimi K2 模型更新,带来更强的代码能力、更快的 API
Kimi K2 又又又提速了
经过工程师们的不懈努力,kimi-k2-turbo-preview 模型输出速度已经提升至每秒 60 Tokens,最高可达每秒 100 Tokens。目前仍然享受 5 折特惠价格(模型每百万 tokens 输入价格(缓存命中)¥2.00,输入价格(缓存未命中)¥8.00,输出价
Qwen-Image-Edit: Image Editing with Higher Quality and Efficiency
We are excited to introduce Qwen-Image-Edit, the image editing version of Qwen-Image. Built upon our 20B Qwen-Image model, Qwen-Image-Edit s
Qwen-Image: Crafting with Native Text Rendering
We are thrilled to release Qwen-Image, a 20B MMDiT image foundation model that achieves significant advances in complex text rendering and p
Qwen-MT: Where Speed Meets Smart Translation
Introduction Here we introduce the latest update of Qwen-MT (qwen-mt-turbo) via Qwen API. This update builds upon the powerful Qwen3, levera
MiniMax-01
We are delighted to introduce two remarkable models, **MiniMax-Text-01** and **MiniMax-VL-01**.
How People Use Claude for Support, Advice, and Companionship
在约 450 万段 Claude.ai 对话中仅 2.9% 为情感类,浪漫与性角色扮演合计不足 0.5%——AI 伴侣比想象中罕见,与 OpenAI/MIT 针对 ChatGPT 的独立研究结论一致。
Time to Speak Some Dialects, Qwen-TTS!
Introduction Here we introduce the latest update of Qwen-TTS (qwen-tts-latest or qwen-tts-2025-05-22) through Qwen API . Trained on a large-
Qwen VLo: From \"Understanding\" the World to \"Depicting\" It
Introduction The evolution of multimodal large models is continually pushing the boundaries of what we believe technology can achieve. From
Anthropic 教育报告:大学生如何使用 Claude
计算机科学学生占 Claude.ai 对话 38.6% 但仅占学位 5.4%,严重超量;商科、健康、人文则明显不足,既反映 Claude 在 STEM 社区的高知名度,也说明其在 STEM 类任务上适配性更强。
Qwen2.5 Omni: See, Hear, Talk, Write, Do It All!
We release Qwen2.5-Omni, the new flagship end-to-end multimodal model in the Qwen series. Designed for comprehensive multimodal perception,
Qwen2.5-VL-32B: Smarter and Lighter
Introduction At the end of January this year, we launched the Qwen2.5-VL series of models, which received widespread attention and positive
QwQ-32B: Embracing the Power of Reinforcement Learning
QWEN CHAT Hugging Face ModelScope DEMO DISCORD
技术报告:Muon 优化器的首次大规模训练实践
近期,基于矩阵正交化(matrix orthogonalization)的 Muon 优化器在小规模语言模型训练中展现出了优异的性能,但其在大模型训练中的可扩展性尚未得到验证。我们发现了两个提升 Muon 可扩展性的关键技术:(1)引入权重衰减(weight decay);(2)
介绍一下 MoBA:面向长文本大模型的混合块注意力机制
MoBA通过将专家混合系统(Mixture of Experts, MoE)的思想与稀疏注意力(sparse attention)相结合,为大语言模型中的长文本处理方式带来革命性变化。
为什么要推出 Kimi Latest 模型?
2024 年 1 月 31 日,Kimi 开放平台开启公测,推出了最高支持 128k 上下文大小的
Qwen2.5-Max: Exploring the Intelligence of Large-scale MoE Model
It is widely recognized that continuously scaling both data size and model size can lead to significant improvements in model intelligence.
Qwen2.5 VL! Qwen2.5 VL! Qwen2.5 VL!
We release Qwen2.5-VL, the new flagship vision-language model of Qwen and also a significant leap from the previous Qwen2-VL. To try the lat
Global-batch load balance almost free lunch to improve your MoE LLM training
Background The Mixture-of-Experts (MoEs) architecture has become a popular model-parameter-scale-up technique. Typically, one MoE layer cons
QwQ: Reflect Deeply on the Boundaries of the Unknown
Note: This is the pronunciation of QwQ: /kwju:/ , similar to the word “quill”.
Qwen2.5-Coder Series: Powerful, Diverse, Practical.
Introduction Today, we are excited to open source the “Powerful”, “Diverse”, and “Practical” Qwen2.5-Coder series, dedicated to continuously
Qwen2.5: A Party of Foundation Models!
Introduction In the past three months since Qwen2’s release, numerous developers have built new models on the Qwen2 language models, providi
Qwen2.5-LLM: Extending the boundary of LLMs
Introduction In this blog, we delve into the details of our latest Qwen2.5 series language models. We have developed a range of decoder-only
Qwen2.5-Coder: Code More, Learn More!
Introduction In early April, we introduced CodeQwen1.5, which garnered significant attention from the community. Since then, we have been wo
使用 Unreal5 游戏引擎和 Kimi 大模型开发交互式游戏
使用 Unreal5 游戏引擎和 Kimi 大模型开发交互式游戏
Qwen2-VL: To See the World More Clearly
After a year’s relentless efforts, today we are thrilled to release Qwen2-VL! Qwen2-VL is the latest version of the vision language models b
Qwen2-Audio: Chat with Your Voice!
To achieve the objective of building an AGI system, the model should be capable of understanding information from different modalities. Than
Introducing Qwen2-Math
🚨 This model mainly supports English. We will release bilingual (English and Chinese) math models soon. Introduction Over the past year, we
Kimi 企业级 API 正式发布
Kimi 企业级 API 正式发布
Kimi 开放平台 Office Hour Season 1 Recap
Kimi 开放平台 Office Hour Season 1 Recap
Hello Qwen2
Introduction After months of efforts, we are pleased to announce the evolution from Qwen1.5 to Qwen2. This time, we bring to you:
Kimi API 还没用起来?请看这篇无门槛快速入门指南
Kimi API 已经发布一个多月了,有没有用它做点有意思的事情?
Kimi 大模型 API 更新了,也期待在「亚马逊云科技中国峰会」见到大家 | 开发者速递
Kimi 大模型 API 更新了,也期待在「亚马逊云科技中国峰会」见到大家 | 开发者速递
Notes on Qwen-Max-0428
Previously, we opensourced a series of Qwen1.5 model ranging from 0.5 to 110 billion parameters. Now, we release a larger model, Qwen-Max-04
Qwen1.5-110B: The First 100B+ Model of the Qwen1.5 Series
Introduction Recently we have witnessed a burst of large-scale models with over 100 billion parameters in the opensource community. These mo
Code with CodeQwen1.5
Introduction The advent of advanced programming tools, which harnesses the power of large language models (LLMs), has significantly enhanced
Qwen1.5-32B: Fitting the Capstone of the Qwen1.5 Language Model Series
Introduction The open-source community has long sought a model that strikes an ideal balance between performance, efficiency, and memory foo
Qwen1.5-MoE: Matching 7B Model Performance with 1/3 Activated Parameters
Introduction Since the surge in interest sparked by Mixtral, research on mixture-of-expert (MoE) models has gained significant momentum. Bot
Research
Our research teams investigate the safety, inner workings, and societal impacts of AI models—so that artificial intelligence has a positive
Introducing Qwen1.5
Introduction In recent months, our focus has been on developing a “good” model while optimizing the developer experience. As we progress tow
Introducing Qwen
4 months after our first release of Qwen-7B, which is the starting point of our opensource journey of large language models (LLM), we now pr
projectdeal
Got some quirky supplies for your next creative project:
OFASys: Enabling Multitask Learning with One Line of Code!
Intro Generalist Models are hot! We all see an opportunity towards a real generalist model by multimodal multitask learning. We previously r
Chinese CLIP: Contrastive Vision-Language Pretraining in Chinese
CLIP1 is a phenomenal playmaker in vision and multimodal representation learning. It plays not only as a foundation model but also a bridge
OFA: Towards Building a One-For-All Model
2022 is a year of generalist models! With the bloom of multimodal pretraining, especially the unified model, we have witnessed the opportuni
Letâs innovate together.
与 Apple 一起构建令人惊叹的机器学习体验。探索 Apple 为开发者和研究人员提供的机遇。
Contact Sales
能为您的组织带来哪些价值吗?请填写以下表格,我们的销售团队成员将尽快与您联系。