OpenAI 开源模型泄露：六大技术细节

华尔街见闻2025-08-02

OpenAI 可能即将发布的开源大模型的详细技术细节来了，以下是根据泄露信息整理模型架构：1200亿参数的混合专家模型（MoE）据爆料，OpenAI 可能会发布两款模型：一款 1200亿（120B）参数的混合专家（MoE）模型：其在推理时仅激活约 50-60亿（5B/6B）参数。这意味着它能在保持巨大知识容量的同时，实现极高的推理效率，大幅降低运行成本一款 200亿（20B）参数的稠密...

网页链接

免责声明：本文观点仅代表作者个人观点，不构成本平台的投资建议，本平台不对文章信息准确性、完整性和及时性做出任何保证，亦不对因使用或信赖文章信息引发的任何损失承担责任。

精彩评论

我们需要你的真知灼见来填补这片空白

发表看法

{"i18n":{"language":"zh_CN"},"isChannel":false,"data":{"share":"https://www.laohu8.com/m/news/2556169842?lang=zh_CN&edition=full","thumbnail":"https://wpimg-wscn.awtmt.com/a5992534-c8ce-4384-9da5-077a59520bf5.png","is_english":false,"pubTime":"2025-08-02 07:37","share_image_url":"https://static.laohu8.com/e9f99090a1c2ed51c021029395664489","id":"2556169842","market":"us","top_or_hot":-1,"title":"OpenAI 开源模型泄露：六大技术细节","media":"华尔街见闻","content":"<div>\n<p>OpenAI 可能即将发布的开源大模型的详细技术细节来了，以下是根据泄露信息整理\n\n模型架构：1200亿参数的混合专家模型（MoE）\n据爆料，OpenAI 可能会发布两款模型：\n一款 1200亿（120B）参数的混合专家（MoE）模型：其在推理时仅激活约 50-60亿（5B/6B）参数。这意味着它能在保持巨大知识容量的同时，实现极高的推理效率，大幅降低运行成本\n一款 200亿（20B）参数的稠密...</p>\n\n<a href=\"https://wallstreetcn.com/articles/3752494\">网页链接</a>\n\n</div>\n","source":"wallstreetcn_hot_news","html":"<!DOCTYPE html>\n<html>\n<head>\n<meta http-equiv=\"Content-Type\" content=\"text/html; charset=utf-8\" />\n<meta name=\"viewport\" content=\"width=device-width,initial-scale=1.0,minimum-scale=1.0,maximum-scale=1.0,user-scalable=no\"/>\n<meta name=\"format-detection\" content=\"telephone=no,email=no,address=no\" />\n<title>OpenAI 开源模型泄露：六大技术细节</title>\n<style type=\"text/css\">\na,abbr,acronym,address,applet,article,aside,audio,b,big,blockquote,body,canvas,caption,center,cite,code,dd,del,details,dfn,div,dl,dt,\nem,embed,fieldset,figcaption,figure,footer,form,h1,h2,h3,h4,h5,h6,header,hgroup,html,i,iframe,img,ins,kbd,label,legend,li,mark,menu,nav,\nobject,ol,output,p,pre,q,ruby,s,samp,section,small,span,strike,strong,sub,summary,sup,table,tbody,td,tfoot,th,thead,time,tr,tt,u,ul,var,video{ font:inherit;margin:0;padding:0;vertical-align:baseline;border:0 }\nbody{ font-size:16px; line-height:1.5; color:#999; background:transparent; }\n.wrapper{ overflow:hidden;word-break:break-all;padding:10px; }\nh1,h2{ font-weight:normal; line-height:1.35; margin-bottom:.6em; }\nh3,h4,h5,h6{ line-height:1.35; margin-bottom:1em; }\nh1{ font-size:24px; }\nh2{ font-size:20px; }\nh3{ font-size:18px; }\nh4{ font-size:16px; }\nh5{ font-size:14px; }\nh6{ font-size:12px; }\np,ul,ol,blockquote,dl,table{ margin:1.2em 0; }\nul,ol{ margin-left:2em; }\nul{ list-style:disc; }\nol{ list-style:decimal; }\nli,li p{ margin:10px 0;}\nimg{ max-width:100%;display:block;margin:0 auto 1em; }\nblockquote{ color:#B5B2B1; border-left:3px solid #aaa; padding:1em; }\nstrong,b{font-weight:bold;}\nem,i{font-style:italic;}\ntable{ width:100%;border-collapse:collapse;border-spacing:1px;margin:1em 0;font-size:.9em; }\nth,td{ padding:5px;text-align:left;border:1px solid #aaa; }\nth{ font-weight:bold;background:#5d5d5d; }\n.symbol-link{font-weight:bold;}\n/* header{ border-bottom:1px solid #494756; } */\n.title{ margin:0 0 8px;line-height:1.3;color:#ddd; }\n.meta {color:#5e5c6d;font-size:13px;margin:0 0 .5em; }\na{text-decoration:none; color:#2a4b87;}\n.meta .head { display: inline-block; overflow: hidden}\n.head .h-thumb { width: 30px; height: 30px; margin: 0; padding: 0; border-radius: 50%; float: left;}\n.head .h-content { margin: 0; padding: 0 0 0 9px; float: left;}\n.head .h-name {font-size: 13px; color: #eee; margin: 0;}\n.head .h-time {font-size: 11px; color: #7E829C; margin: 0;line-height: 11px;}\n.small {font-size: 12.5px; display: inline-block; transform: scale(0.9); -webkit-transform: scale(0.9); transform-origin: left; -webkit-transform-origin: left;}\n.smaller {font-size: 12.5px; display: inline-block; transform: scale(0.8); -webkit-transform: scale(0.8); transform-origin: left; -webkit-transform-origin: left;}\n.bt-text {font-size: 12px;margin: 1.5em 0 0 0}\n.bt-text p {margin: 0}\n</style>\n</head>\n<body>\n<div class=\"wrapper\">\n<header>\n<h2 class=\"title\">\nOpenAI 开源模型泄露：六大技术细节\n</h2>\n\n<h4 class=\"meta\">\n\n\n2025-08-02 07:37 北京时间&nbsp;&nbsp;&nbsp;<a href=https://wallstreetcn.com/articles/3752494><strong>华尔街见闻</strong></a>\n\n\n</h4>\n\n</header>\n<article>\n<div>\n<p>OpenAI 可能即将发布的开源大模型的详细技术细节来了，以下是根据泄露信息整理\n\n模型架构：1200亿参数的混合专家模型（MoE）\n据爆料，OpenAI 可能会发布两款模型：\n一款 1200亿（120B）参数的混合专家（MoE）模型：其在推理时仅激活约 50-60亿（5B/6B）参数。这意味着它能在保持巨大知识容量的同时，实现极高的推理效率，大幅降低运行成本\n一款 200亿（20B）参数的稠密...</p>\n\n<a href=\"https://wallstreetcn.com/articles/3752494\">网页链接</a>\n\n</div>\n\n\n</article>\n</div>\n</body>\n</html>\n","isBrief":false,"type":0,"news_type":1,"symbol":"LU2028103732.USD","symbol_name":"ALLIANZ GLOBAL SUSTAINABILITY \"AMG\" (USD) INC","start_time":0,"source_url":"https://wallstreetcn.com/articles/3752494","article_id":"2556169842","we_media_id":null,"thumbnails":["https://wpimg-wscn.awtmt.com/a5992534-c8ce-4384-9da5-077a59520bf5.png"],"rights":null,"url":"https://stock-news.laohu8.com/highlight/detail?id=2556169842","pubTimestamp":1754091459,"columns":[],"sourceInfo":{"source_id":"wallstreetcn_hot_news","name":"华尔街见闻"},"weMediaInfo":null,"summary":"OpenAI即将发布两款开源大模型的六大技术特点包括：采用Float4量化配合英伟达Blackwell芯片训练；使用带范围限制的SwiGLU激活函数；通过YaRN技术将上下文窗口从4K扩展至128K；采用滑动窗口注意力和注意力汇聚机制处理长文本；底层架构融合Llama/Mixtral特点并广泛使用偏置项。","collect":0,"end_time":0,"defaultTopTitle":"wallstreetcn.com","property":[],"viewcount":null,"language":"zh","relate_stocks":{"LU2028103732.USD":"ALLIANZ GLOBAL SUSTAINABILITY \"AMG\" (USD) INC","LU0672654240.SGD":"FTIF - Franklin US Opportunities A Acc SGD-H1","LU0444971666.USD":"天利全球科技基金","LU0971096721.USD":"富达环球金融服务 A","LU1429558221.USD":"Natixis Loomis Sayles US Growth Equity RA USD","LU2764262908.HKD":"BGF GLOBAL UNCONSTRAINED EQUITY \"A2\" (HKD) ACC","LU1116320737.USD":"BGF SYSTEMATIC GLOBAL ENHANCED EQUITY YIELD \"A6\" (USD) INC","LU0096362180.USD":"CT (LUX) I GLOBAL FOCUS \"DU\" (USD)","LU0708995401.HKD":"FRANKLIN U.S. OPPORTUNITIES \"A\" (HKD) ACC","LU0342679015.USD":"ALLIANZ GLOBAL EQUITY UNCONSTRAINED \"AT\" (USD) ACC","IE00B19Z8W00.USD":"FTGF CLEARBRIDGE US LARGE CAP GROWTH \"A\" INC","LU2471134952.CNY":"INVESCO GLOBAL EQUITY INCOME ADVANTAGE \"A\" (CNYHDG) INC","LU0256863902.USD":"ALLIANZ US EQUITY \"AT\" (USD) ACC","LU0154236417.USD":"BGF US FLEXIBLE EQUITY \"A2\" ACC","LU0079474960.USD":"联博美国增长基金A","LU2023251221.USD":"ALLIANZ GLOBAL SUSTAINABILITY \"AM\" (USD) INC","LU0256863811.USD":"ALLIANZ US EQUITY \"A\" INC","SG9999014880.SGD":"大华全球优质成长基金Acc SGD","LU0289739343.SGD":"SUSTAINABLE GLOBAL THEMATIC PORTFOLIO \"A\" (SGD) ACC","LU0289941410.SGD":"AB FCP I Dynamic Diversified AX SGD","LU1815336760.USD":"THREADNEEDLE (LUX) GLOBAL TECHNOLOGY \"AUP\" (USD) INC","LU0215105999.USD":"SCHRODER ISF GLOBAL EQUITY \"A\" ACC","LU0823421416.USD":"BNP PARIBAS DISRUPTIVE TECHNOLOGY \"C\" (USD) INC","LU0106261372.USD":"SCHRODER ISF US LARGE CAP \"A\" ACC","LU2326559502.SGD":"Natixis Loomis Sayles US Growth Equity P/A SGD-H","LU0158827948.USD":"ALLIANZ GLOBAL SUSTAINABILITY \"A\" (USD) INC","LU2089284900.SGD":"Allianz Global Sustainability Cl AM Dis H2-SGD","IE00BQXX3C00.GBP":"GUINNESS GLOBAL INNOVATORS \"C\" (GBP) ACC","LU2720916845.USD":"BGF GLOBAL UNCONSTRAINED EQUITY \"A2\" (USD) ACC","LU0289960550.SGD":"AB FCP I - GLOBAL EQUITY BLEND PORTFOLIO 'A' (SGD) ACC","LU0234572021.USD":"高盛美国核心股票组合Acc","LU2404859667.USD":"FUNDSMITH EQUITY FUND \"R\" (USD) ACC","LU2746668461.USD":"MANULIFE DYNAMIC LEADERS \"AA\" (USD) ACC","LU2125154778.USD":"ALLSPRING GLOBAL EQUITY ENHANCED INCOME \"A\" (USD) INC","LU1043141123.HKD":"NINETY ONE GSF GLOBAL FRANCHISE \"A\" (HKD) INC 2","LU2111349929.HKD":"ALLIANZ GLOBAL SUSTAINABILITY \"AM\" (HKD) INC","IE00BZ1G4Q59.USD":"LEGG MASON CLEARBRIDGE US EQUITY SUSTAINABILITY LEADER \"A\"(USD) INC (A)","LU0690374615.EUR":"FUNDSMITH EQUITY \"R\" (EUR) ACC","LU2244417387.USD":"FIDELITY SUSTAINABLE US EQUITY \"A\" (USD) ACC","LU1035775433.USD":"AB SICAV I - AMERICAN GROWTH PORTFOLIO \"AD\" (USD) INC","SG9999017495.SGD":"UGDP UNITED GLOBAL QUALITY GROWTH \"B\" (SGD) ACC","LU0109391861.USD":"富兰克林美国机遇基金A Acc","LU0994945656.USD":"NINETY ONE GSF GLOBAL FRANCHISE \"A\" (USD) INC 2","LU1934455277.USD":"AB SICAV I LOW VOLATILITY TOTAL RETURN EQUITY PORT \"AD\" (USD) INC","LU1814569148.SGD":"WELLINGTON GLOBAL QUALITY GROWTH \"D\" (SGDHDG) ACC","LU0353189680.USD":"富国美国全盘成长基金Cl A Acc","BK4585":"ETF&股票定投概念","SG9999018857.SGD":"United Global Quality Growth Fd Cl Acc SGD-H"},"translate_title":"OpenAI open source model leaked: Six major technical details","themeId":null,"isJumpTheme":false,"ttsUrl":null,"symbols_score_info":{"K":1},"content_text":"OpenAI 可能即将发布的开源大模型的详细技术细节来了，以下是根据泄露信息整理\n\n模型架构：1200亿参数的混合专家模型（MoE）\n据爆料，OpenAI 可能会发布两款模型：\n一款 1200亿（120B）参数的混合专家（MoE）模型：其在推理时仅激活约 50-60亿（5B/6B）参数。这意味着它能在保持巨大知识容量的同时，实现极高的推理效率，大幅降低运行成本\n一款 200亿（20B）参数的稠密模型：作为更小巧、更易于部署的版本\n目前来看，这两款模型将专注于文本处理，暂时不涉及多模态能力\n训练技术：或采用 Float4 与 英伟达最新Blackwell 芯片\n为了极致的效率，模型可能采用了 Float4进行训练或量化。这是一种非常激进的量化方案，可以极大地压缩模型体积并提升运算速度\n据推测，这可能是借助英伟达最新发布的 Blackwell 架构 GPU 完成的，因为该系列芯片原生支持 Float4 运算。另一种可能性是，模型在训练后通过 训练后量化（PTQ） 技术压缩到了 Float4\n激活函数：带范围限制的 SwiGLU\n为了配合 Float4 量化，模型可能采用了 SwiGLU 激活函数，并对其输出范围进行了 裁剪，限制在 -7 到 7 之间\n这类似于经典的 ReLU6 函数，其目的是为了消除激活值中的极端异常值，确保数值分布更稳定，从而降低量化过程中带来的精度损失。这对于 Float4 这种低精度格式至关重要\n上下文窗口：通过 YaRN 技术扩展至 128K\n模型将拥有128K 的超长上下文窗口，但它并非从头开始训练的。据推测，该模型的基础上下文窗口为 4K，之后在训练中途采用 YaRN等技术将其无缝扩展至 128K\n注意力机制：滑动窗口注意力（SWA）与注意力汇聚（Attention Sinks）\n为了高效处理 128K 的长文本，模型采用了两大关键技术：\n滑动窗口注意力（SWA）：窗口大小为 128。这意味着在计算注意力时，每个词元（token）只需关注其邻近的 128 个词元，从而将计算复杂度从二次方降低到线性级别\n注意力汇聚（Attention Sinks）：为了解决 SWA 会遗忘早期重要信息的问题，模型引入了注意力汇聚技术。该技术强制模型始终关注最开始的几个（例如 4 或 8 个）关键 token，确保模型在处理长序列时不会失忆。NVIDIA 的 TensorRT-LLM 也已支持此功能\n底层架构：融合 Llama/Mixtral 特点并使用偏置项\n模型的基础架构很可能借鉴了 Llama 和 Mixtral 等成功的开源模型。关键特征包括：\n合并的 QKV 矩阵：将注意力机制中的查询（Q）、键（K）、值（V）矩阵合并，以优化计算效率\n广泛使用偏置项（Biases）：与一些模型（如 Llama）去掉偏置项的做法不同，该模型似乎在所有模块（包括 MLP、注意力层甚至 MoE 的路由层）都保留了偏置项，这可能有助于提升模型的拟合能力\n来源：AI寒武纪，原文标题：《OpenAI 开源模型泄露：六大技术细节》风险提示及免责条款\n\n            市场有风险，投资需谨慎。本文不构成个人投资建议，也未考虑到个别用户特殊的投资目标、财务状况或需要。用户应考虑本文中的任何意见、观点或结论是否符合其特定状况。据此投资，责任自负。","kind":"news","is_publish_news":true,"is_publish_highlight":false,"is_publish_live":false,"is_publish_wemedia":null,"editions":null,"column":"","sentiment":"0","news_tag":"","news_rank":0,"isVideo":false,"video":null,"symbols":[],"gpt_button":0,"need_auth":false,"need_login_tip":false,"code":"91000000","status":"200"},"commentList":[],"isCommentEnd":true,"newsSizeData":{"likeSize":0,"commentSize":0,"repostSize":0,"favoriteSize":0,"likeStatus":false,"favoriteStatus":false},"APP":{"userAgent":"Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; ClaudeBot/1.0; +claudebot@anthropic.com)","isDev":false,"isTTM":false,"tenantId":"TBCN","deviceId":"web-server-community-laohu8-v3","version":"4.43.1","shortVersion":"4.43.1","platform":"web","vendor":"web","appName":"laohu8","isIOS":false,"isAndroid":false,"isTiger":false,"isTHS":false,"isWeiXin":false,"isWeiXinMini":false,"isWeiBo":false,"isQQ":false,"isBaiduSwan":false,"isBaiduBox":false,"isDingTalk":false,"isToutiao":false,"isOnePlus":false,"isHuaWei":false,"isXiaomi":false,"isXiaomiWebView":false,"isOppo":false,"isVivo":false,"isSamsung":false,"isMobile":false},"href":"/m/news/2556169842","isCrawlerRequest":true}