豆包1.5·深度思考模型

Search documents
全网都在等梁文锋
凤凰网财经· 2025-04-29 12:39
以下文章来源于凤凰网科技 ,作者凤凰网科技 凤凰网科技 . 凤凰科技频道官方账号,带你直击真相。 来源|凤凰网科技 作者|姜凡 编辑|董雨晴 五月将至,中美科技巨头或将迎来新一轮巅峰对决。 先是在4月中旬,OpenAI一口气发布了GPT-4.1 o3、o4 mini系列模型;谷歌则拿出了Gemini 2.5 Flash Preview,一个混合推理模型;与谷歌同 一天,豆包在杭州巡展中正式发布了1.5·深度思考模型,在多模态上展现出了更强的实力。凤凰网科技从行业人士处了解到,阿里的下一代大模型 Qwen3也将于本月内发布。 混战之下,那股"神秘的东方力量"似乎也在悄悄准备着新的发布。 敏感的神经之下,一点蛛丝马迹都会被放大。 昨日,全球最大AI开源社区Hugging Face首席执行官Clément Delangue在社交平台发布了一条耐人 寻味的动态。这条动态仅由三个眼睛的表情符号构成,并附上了DeepSeek团队在Hugging Face平台的官方资源库入口。 这组充满悬念的组合引发科技圈热议,业内普遍推测DeepSeek R2模型已进入发布倒计时。 01 DeepSeek R2发布已进入倒计时? 近半个 ...
AI模型迭代潮涌,港股通科技ETF(513860)高开高走涨超1%
Sou Hu Cai Jing· 2025-04-25 02:49
截至2025年4月25日 10:04,中证港股通科技指数(931573)上涨0.32%,成分股高伟电子(01415)上涨4.75%,晶泰控股(02228)上涨3.67%,联想集团(00992)上涨 3.51%,零跑汽车(09863)上涨3.38%,信义光能(00968)上涨3.10%。港股通科技ETF(513860)上涨1.03%,最新价报0.69元。拉长时间看,截至2025年4月24 日,港股通科技ETF近1周累计上涨5.44%,涨幅排名可比基金1/6。 流动性方面,港股通科技ETF盘中换手3.09%,成交5209.36万元。拉长时间看,截至4月24日,港股通科技ETF近1周日均成交1.99亿元。 消息面上,AI模型及应用持续迭代升级。近期,腾讯微信集成元宝AI功能,可解析公众号文章和图文,提升用户体验。OpenAI发布GPT-4.1,o3和o4-mini。 火山引擎发布豆包1.5·深度思考模型,升级豆包·文生图模型3.0、豆包·视觉理解模型;谷歌发布首款混合推理模型Gemini2.5Flash,性能一举击败Claude3.7, 比肩o4-mini。而且,关闭思考模式成本直降600%。 规模方面,港股通科技 ...
【数字经济周报】美国限制英伟达H20 GPU、AMDMI308等AI芯片对华出口-20250421
GUOTAI HAITONG SECURITIES· 2025-04-21 14:37
Investment Rating - The report does not explicitly state an investment rating for the industry Core Insights - The report highlights significant developments in the technology sector, particularly in semiconductor, automotive electronics, AI, and metaverse industries, indicating a dynamic investment landscape Summary by Sections 1. Weekly Capital Market Overview - During the period from April 12 to April 18, 2025, there were 14 financing events in the technology sector, all occurring domestically, with advanced manufacturing, AI, and VR/AR leading the financing activity with 9, 4, and 1 events respectively [9] - Notable IPOs included Hong Kong's Everbright Digital and China's Hong Kong-listed Kent Catalysts, raising significant capital for their respective operations [12][16] 2. Semiconductor Sector Dynamics - Intel is nearing the sale of 51% of its Altera chip division to Silver Lake for $8.75 billion, a significant reduction from its previous acquisition cost [35] - The U.S. has imposed new export restrictions on NVIDIA's H20 and AMD's MI308 AI chips to China, potentially impacting NVIDIA with a projected loss of $5.5 billion [38] - AMD announced its first 2nm chip, the core complex die (CCD), which is set to enhance its product offerings in the high-performance computing market [39] 3. Automotive Electronics Sector Dynamics - Onsemi has canceled its $6.9 billion acquisition of Allegro MicroSystems, which was aimed at strengthening its position in the automotive chip market [41] - Yika Technology successfully validated its single-chip "cabin driving parking integration" solution, demonstrating a cost reduction of at least 20% compared to traditional methods [42] - SAIC and Huawei launched a new brand "SAIC Shangjie," marking a significant collaboration in the smart vehicle sector [43] 4. AI Sector Dynamics - OpenAI and Google have made advancements in AI models, with OpenAI releasing o3/o4-mini and Google unveiling a dolphin language model [4] - The AI sector continues to see rapid growth, with significant investments and product launches from leading companies [4] 5. Metaverse Sector Dynamics - Huawei introduced new smart glasses, and partnerships in the metaverse space are expanding, indicating a growing interest in immersive technologies [5]
豆包1.5深度思考模型发布:暴砍参数量,能看图思考,数学编程超DeepSeek-R1
3 6 Ke· 2025-04-17 08:54
智东西4月17日报道,今天,火山引擎正式发布了豆包1.5·深度思考模型,采用MoE架构,总参数量为200B,激活参数仅20B,在多项基准测试 中达到或接近全球第一梯队水平。 在数学、编程、科学领域,豆包1.5·深度思考模型的测试表现超过DeepSeek-R1、QwQ-32B等国产推理模型,与OpenAI o1、o3-mini-high等模 型能力相仿。在高难度通用人工智能测试ARC-AGI上,其得分更是超过了OpenAI o1和o3-mini-high。 豆包1.5·深度思考模型具有"边想边搜"、"视觉理解"等实用能力,即日起,企业用户可以在火山方舟平台使用豆包1.5·深度思考模型。火山引擎 还升级了豆包·文生图模型3.0、豆包·视觉理解模型。 数据显示,截至2025年3月底,豆包大模型日均tokens调用量已超过12.7万亿,是2024年12月的3倍,是一年前刚刚发布时的106倍。IDC报告显 示,2024年中国公有云大模型调用量激增,火山引擎以46.4%的市场份额位居中国市场第一。 01.高并发场景延迟仅20毫秒,可边想边搜、视觉理解 豆包1.5·深度思考模型较小的参数量和激活参数量,使其具备更低的训练和 ...