Workflow
TRAE
icon
Search documents
字节跳动豆包大模型2.0发布,多数基准达SOTA水平
Sou Hu Cai Jing· 2026-02-14 15:57
豆包 2.0 全面升级了多模态能力,在各类视觉理解任务上均达到世界顶尖水平,视觉推理、感知能力、空间推理与长上下文理解能力表现尤为突出,豆包 2.0 Pro 在大多数相关基准测试中取得最高分。 面对动态场景,豆包 2.0 强化了对时间序列与运动感知的理解能力,在TVBench等关键测评中处于领先位置,且在 EgoTempo 基准上超过了人类分数,表 明它对"变化、动作、节奏"这类信息的捕捉更为稳定,在工程侧可用性更高。 长视频场景中,豆包 2.0 在大多评测上超越了其他顶尖模型,且在多个流式实时问答视频基准测试中表现优异,能作为 AI 助手完成实时视频流分析、环 境感知、主动纠错与情感陪伴,实现从被动问答到主动指导的交互升级,可应用于健身、穿搭等陪伴场景。 LLM与 Agent 表现大幅强化,长程任务执行能力提升 IT之家 2 月 14 日消息,字节跳动宣布,今天,豆包大模型正式进入 2.0 阶段。豆包 2.0(Doubao-Seed-2.0)围绕大规模生产环境下的使用需求做了系统性 优化,依托高效推理、多模态理解与复杂指令执行能力,更好地完成真实世界复杂任务。 IT之家注意到,豆包 2.0 系列包含 Pro ...
字节豆包2.0发布:推理成本降一个数量级,正面对标GPT-5和Gemini 3
硬AI· 2026-02-14 11:37
分析认为,在现实世界复杂任务中, 由于大规模推理与长链路生成将消耗大量token,豆包2.0的成本优 势将成为关键竞争力 。这标志着字节跳动在大模型商业化应用上迈出重要一步。 01 多模态能力达到世界顶尖水平 豆包2.0全面升级了多模态能力,在视觉推理、感知能力、空间推理与长上下文理解等任务上表现突出。 字节发布豆包2.0,旗舰版Pro全面对标GPT-5.2与Gemini 3 Pro。新模型在多模态、数学及编程等领域达到业界顶尖, 同时将推理成本降低约一个数量级,显著提升Agent应用性价比。目前已接入豆包App、TRAE及火山引擎API。 硬·AI 作者 | 董 静 编辑 | 硬 AI 字节跳动旗下豆包大模型正式进入2.0阶段,推出面向Agent时代的系统性升级版本。 新版本在保持与 GPT-5.2和Gemini 3 Pro相当性能的同时,将推理成本降低约一个数量级 ,为大规模生产环境下的复杂任 务执行提供更具竞争力的解决方案。 2月14日,字节跳动宣布,豆包2.0系列包含Pro、Lite、Mini三款通用Agent模型和专门的Code模型。 其 中旗舰版豆包2.0 Pro全面对标GPT-5.2与Gemin ...
豆包再扔王炸!2.0发布:推理成本降一个数量级,正面对标GPT-5和Gemini 3
华尔街见闻· 2026-02-14 10:53
字节跳动旗下豆包大模型正式进入2.0阶段,推出面向Agent时代的系统性升级版本。 新版本在保持与GPT-5.2和Gemini 3 Pro相当性能的同时,将推理成本降 低约一个数量级 ,为大规模生产环境下的复杂任务执行提供更具竞争力的解决方案。 2月14日,字节跳动宣布,豆包2.0系列包含Pro、Lite、Mini三款通用Agent模型和专门的Code模型。 其中旗舰版豆包2.0 Pro全面对标GPT-5.2与Gemini 3 Pro,在多数视觉理解基准测试中达到业界最高水平,并在数学奥赛IMO、CMO和编程竞赛ICPC中获得金牌成绩。 该系列模型已全面上线。豆包2.0 Pro已接入豆包App、电脑端和网页版的"专家"模式,Code版本已集成至AI编程产品TRAE,火山引擎同步上线面向企业和开 发者的API服务。 分析认为,在现实世界复杂任务中, 由于大规模推理与长链路生成将消耗大量token,豆包2.0的成本优势将成为关键竞争力 。这标志着字节跳动在大模型商 业化应用上迈出重要一步。 多模态能力达到世界顶尖水平 豆包2.0全面升级了多模态能力,在视觉推理、感知能力、空间推理与长上下文理解等任务上表现突出。 ...
字节豆包2.0发布:推理成本降一个数量级,正面对标GPT-5和Gemini 3
Hua Er Jie Jian Wen· 2026-02-14 09:29
Core Insights - ByteDance's Doubao model has officially entered its 2.0 phase, offering a systematic upgrade that maintains performance comparable to GPT-5.2 and Gemini 3 Pro while reducing reasoning costs by approximately an order of magnitude, making it a competitive solution for complex tasks in large-scale production environments [1][7] Model Features - The Doubao 2.0 series includes three general-purpose agent models (Pro, Lite, Mini) and a specialized Code model, with the flagship Doubao 2.0 Pro achieving top scores in visual understanding benchmarks and winning gold medals in mathematics and programming competitions [1][5] - Doubao 2.0 has significantly upgraded its multimodal capabilities, excelling in visual reasoning, perception, spatial reasoning, and long-context understanding tasks [2] Performance Metrics - In dynamic scene understanding, Doubao 2.0 leads in key assessments like TVBench and surpasses human scores in EgoTempo, demonstrating stable capture of information related to changes, actions, and rhythms [4] - The model outperforms other leading models in long video scenarios and excels in real-time video question-answering benchmarks, enabling it to function as an AI assistant for real-time video stream analysis and proactive guidance [4] Cost Efficiency - Doubao 2.0 Pro has surpassed GPT-5.2 in SuperGPQA and achieved first place in HealthBench, with overall performance in scientific fields comparable to Gemini 3 Pro and GPT-5.2 [5] - The model's token pricing has been reduced by approximately an order of magnitude, enhancing its competitive edge in large-scale reasoning and long-chain generation scenarios [7] Application and Integration - The Doubao 2.0 Code model has been optimized for programming scenarios, improving code library interpretation and application generation capabilities, and is integrated into the TRAE product [8] - Developers can create interactive projects with minimal prompts, showcasing the model's efficiency in generating complex applications [8] - Doubao 2.0 Pro is now available to end-users through the Doubao App and web platforms, while API services for enterprises and developers have been launched via Volcano Engine [8]
整整21个月,豆包大模型正式进入2.0时代!
量子位· 2026-02-14 08:13
这是 时隔21个月 以来的最大版本的更新。 金磊 发自 凹非寺 量子位 | 公众号 QbitAI 在 Seedance 2.0 和 Seedream 5.0 Lite ,一波接一波爆火之后,豆包把完全体拿出来了—— 豆包大模型2.0 。 像Seedance 2.0已经成为全民玩转的AI,我们也试着做了一个视频: 短短5秒钟,效果确实是足够逼真。 也难怪老外也开始研究怎么注册中国手机号来体验了…… 再如 Seedream 5.0 Lite ,首次支持联网检索,生成的图片也达到了商业化的水平: 而就在今天,在视觉模型火爆之后,豆包终于把那个最核心的大脑拿出来了—— 豆包大模型2.0 。 整体来看,这次豆包大模型2.0在多模态理解、企业级Agent、推理和代码能力上都有了不少的提升: 更直观的提升,体现在榜单测评中。 例如在MathVista、MathVision、MathKangaroo、MathCanvas等数学推理基准上达到业界最优水平。同时,在 LogicVista、VisuLogic 等视觉解谜与逻辑推理基准上,Seed2.0 Pro得分较Seed1.8显著提升。 更强多模态理解:在多模态感知、高精度文字 ...
字节跳动:豆包大模型2.0正式发布
Xin Lang Cai Jing· 2026-02-14 06:29
新浪科技讯 2月14日下午消息,今日,字节跳动豆包大模型2.0正式发布。豆包2.0系列包含Pro、Lite、 Mini三款通用Agent模型和Code模型,灵活适配各类业务场景。 豆包2.0 Pro面向深度推理与长链路任务执行场景,全面对标GPT 5.2与Gemini 3 Pro;2.0 Lite兼顾性能与 成本,综合能力超越上一代主力模型豆包1.8;2.0 Mini面向低时延、高并发与成本敏感场景;Code版 (Doubao-Seed-2.0-Code)专为编程场景打造,与TRAE结合使用效果更佳。 豆包大模型2.0发布 激发创造丰富生活 字节跳动 2026年2月14日 13:54 北京 60 3人 Ind 学节跳动 今天,豆包大模型正式进入2.0阶段。 随着Agent时代到来,大模型将在现实 世 界 发 挥 更 大 作 用 。 豆 包 2.0 (Doubao-Seed-2.0 ) 围绕大规模生 产环境下的使用需求做了系统性优化, 依托高效推理、多模态理解与复杂指令 执行能力,更好地完成真实世界复杂任 务。 张伊茗(萨摩耶金服) ♡ 6 3 r 字 ... 十天注 375 238 989 170个 1 豆包 ...
年度AI产品十大赛道TOP 3|量子位智库AI 100
量子位· 2026-01-31 07:30
Core Insights - The article discusses the significant evolution of AI products in 2025, highlighting a shift from merely "talking" to "doing" [3][4] - The focus is on the transformation of interaction paradigms and the integration of AI into both digital and physical realms [5][6] - The article introduces the "AI 100" product list, categorizing AI products into flagship and innovative segments, along with five major application categories [6][9] Group 1: AI Product Development - AI products have shown differentiated growth across various sectors, with strong demand in general scenarios and AI efficiency, while AI life products are exploring better user experiences [14] - The common goal across all sectors is moving towards end-to-end delivery of productivity, shifting the value measurement from "how well it answers" to "how completely it delivers" [14][15] Group 2: Flagship AI Products - The "Flagship AI 100" and "Innovative AI 100" categories represent the strongest and most promising AI products, respectively [7][13] - The article outlines ten core tracks for AI applications, including AI smart assistants, AI agents, AI browsers, AI workstations, Vibe Coding, AI education, AI entertainment, AI health, multimodal creation, and AI consumer hardware [9][10] Group 3: AI Smart Assistants - AI smart assistants are the most traffic-intensive and revenue-near segment, evolving from answering questions to solving problems [16] - Top products in this category include: - Doubao from ByteDance, with over 57 million daily active users [18] - DeepSeek, known for its innovative interaction method that showcases AI reasoning [20] - Tencent Yuanbao, integrating various social networks for enhanced user experience [22] Group 4: AI Agents - AI agents have transitioned from mere conversational tools to executing tasks [23] - Notable products include: - Nano AI from 360 Group, which integrates over 80 large models for task execution [24] - Kouzi, a one-stop AI office space from ByteDance, automating complex workflows [26] - Xingliu, a new generation AI creation tool from Singularity Star, facilitating end-to-end creative processes [30] Group 5: AI Browsers - AI browsers are evolving from passive information displays to active task executors [32] - Key products include: - QQ Browser from Tencent, which integrates AI capabilities to understand user intent [33] - Quark from Alibaba, combining search, reading, and creation functionalities [36] - Fellou, focusing on a unified search and task experience [40] Group 6: AI Workstations - The competition in AI workstations has shifted from the number of features to complete workflow integration [41] - Leading products include: - Baidu Wenku, transforming from a document tool to a knowledge productivity platform [42] - Feishu, integrating AI capabilities into team workflows [46] - Tiangong, focusing on enhancing office and creative efficiency [50] Group 7: AI Education - AI education products are evolving to provide personalized tutoring and enhance learning experiences [61] - Top products include: - KuaiDui AI from Zuoyebang, focusing on personalized tutoring [62] - XiaoYuan AI from Yuanfudao, assisting parents and teachers in managing homework [65] - CapWords, an innovative language learning tool [69] Group 8: AI Entertainment - AI entertainment products are exploring how to provide unique value beyond traditional non-AI products [70] - Notable products include: - Kapi Camera, which enhances user photography experiences [73] - Xingye, a platform for emotional companionship and content creation [76] - DouDou Game Partner, focusing on gaming companionship [79] Group 9: AI Health - The AI health sector is cautiously exploring compliance and user experience [80] - Key products include: - Antifufu, a health management assistant from Ant Group [81] - XiaoHe AI Doctor, providing health consultations based on authoritative medical data [85] - OtterLife, a gamified health management product [88] Group 10: Multimodal Creation - AI creation tools are becoming integral to daily workflows for content creators [90] - Leading products include: - Jidream AI, focusing on video creation processes [91] - Liblib AI, a comprehensive AI creation platform [95] - Keling AI, a creative productivity platform leveraging short video and advertising [97] Group 11: AI Consumer Hardware - The AI consumer hardware sector is characterized by rapid innovation and high turnover [98] - Notable products include: - Plaud Note, an AI note-taking tool [99] - Thunderbird V3 AI glasses, integrating various functionalities [102] - CocoMate, an emotional companion toy [107]
Node.js之父:手写代码已死
3 6 Ke· 2026-01-21 11:08
Core Viewpoint - The era of human-written code is coming to an end, as AI is fundamentally changing programming practices and roles within the industry [1][4][14]. Group 1: Key Figures and Contributions - Ryan Dahl, the creator of Node.js, emphasized that the age of human coding is over, having previously revolutionized backend development with his framework [3][4]. - Salvatore Sanfilippo, co-founder of Redis, highlighted that programming has been permanently altered by AI, marking a significant shift in the industry [4][5]. - The AI programming tool Copilot, based on OpenAI Codex, has reportedly accelerated development speed by over 50% [8]. Group 2: AI Programming Trends - AI programming and concepts like Vibe Coding have gained significant traction, with tools like Claude Code enabling full-stack development and optimization [8][9]. - ByteDance's native programming tool TRAE generated 100 billion lines of code in 2025, equivalent to the output of 3 million programmers working continuously for a year [10]. - A Stack Overflow report indicated that 84% of developers use AI tools, with 69% believing these tools enhance productivity [10]. Group 3: Future of Programming Roles - The programming landscape is shifting from syntax-focused coding to intent-driven development, where human roles are evolving from code writers to requirement editors [7][20]. - Despite the rise of AI, industry leaders assert that programmers will not be replaced but will instead focus on maintaining and improving AI-generated code [16][20]. - Linus Torvalds, initially critical of AI-generated code, acknowledged its potential as a valuable entry point for new programmers, reinforcing the idea that human oversight remains essential [18][20].
Node.js之父:手写代码已死
量子位· 2026-01-21 10:00
Core Viewpoint - The era of human-written code is coming to an end, as AI programming tools are increasingly taking over coding tasks, fundamentally changing the programming landscape [1][28]. Group 1: Influential Figures and Their Statements - Ryan Dahl, the creator of Node.js, stated that the era of human coding is over, which garnered significant attention with over four million views [2][4]. - Salvatore Sanfilippo, the creator of Redis, echoed this sentiment by asserting that programming has been permanently altered by AI [7][8]. - Linus Torvalds, initially critical of AI-generated code, has shifted his stance, acknowledging the effectiveness of AI in coding while emphasizing that programmers will still be needed for maintenance and oversight [30][34]. Group 2: AI Programming Tools and Their Impact - AI programming tools like OpenAI Codex's Copilot have accelerated development speed by over 50% [15]. - Companies are increasingly adopting AI tools for development, with ByteDance's TRAE generating 100 billion lines of code in 2025, equivalent to the output of 3 million programmers working continuously for a year [22][23]. - A Stack Overflow report indicated that 84% of developers use AI tools, with 69% believing these tools enhance productivity [24]. Group 3: Future Trends and Predictions - Gartner predicts that by 2030, over 80% of enterprises will deeply integrate AI for coding tasks [26]. - The demand for programmers is evolving, with companies now seeking candidates proficient in AI programming tools [28]. - The shift in programming focus is moving from syntax to intent, indicating a transformation in how coding is approached in the AI era [12].
慢雾余弦:VS Code 系 IDE 自动执行 tasks 存在安全风险
Xin Lang Cai Jing· 2026-01-18 04:03
Core Viewpoint - The article highlights a potential security risk associated with IDEs based on VS Code, including Cursor, VS Code, Antigravity, and TRAE, which may automatically execute tasks, potentially triggering malicious code when opening directories [1] Group 1: Security Risks - Slow Fog's Yu Xian warns users about the risk of automatic task execution in VS Code-based IDEs [1] - Users are advised to disable the "automatic task running" feature to prevent malicious code execution [1] - Suggested security measures include setting task.allowAutomaticTasks to off and enabling Workspace Trust in Cursor for risk confirmation when opening new projects [1] Group 2: Mitigation Strategies - The article recommends confirming risks even when choosing to trust the workspace to avoid automatic execution of commands hidden in .vscode/tasks.json [1]