{"id":40564,"date":"2025-08-01T11:20:51","date_gmt":"2025-08-01T03:20:51","guid":{"rendered":"https:\/\/www.1ai.net\/?p=40564"},"modified":"2025-08-01T11:20:51","modified_gmt":"2025-08-01T03:20:51","slug":"%e9%98%b6%e8%b7%83%e6%98%9f%e8%be%b0%e6%96%b0%e4%b8%80%e4%bb%a3%e5%9f%ba%e7%a1%80%e5%a4%a7%e6%a8%a1%e5%9e%8b-step-3-%e6%ad%a3%e5%bc%8f%e5%bc%80%e6%ba%90%ef%bc%9a%e6%8b%a5%e6%9c%89%e5%bc%ba%e5%a4%a7","status":"publish","type":"post","link":"https:\/\/www.1ai.net\/en\/40564.html","title":{"rendered":"Step Star's new generation of basic big model Step 3 is officially open-sourced: with powerful visual perception and complex reasoning ability"},"content":{"rendered":"<p>8 \u6708 1 \u65e5\u6d88\u606f\uff0c<a href=\"https:\/\/www.1ai.net\/en\/tag\/%e9%98%b6%e8%b7%83%e6%98%9f%e8%be%b0\" title=\"\u3010\u67e5\u770b\u542b\u6709[\u9636\u8dc3\u661f\u8fb0]\u6807\u7b7e\u7684\u6587\u7ae0\u3011\" target=\"_blank\" rel=\"noopener\">\u9636\u8dc3\u661f\u8fb0<\/a>\u5ba3\u5e03\u65b0\u4e00\u4ee3<a href=\"https:\/\/www.1ai.net\/en\/tag\/%e5%9f%ba%e7%a1%80%e5%a4%a7%e6%a8%a1%e5%9e%8b\" title=\"\u3010\u67e5\u770b\u542b\u6709[\u57fa\u7840\u5927\u6a21\u578b]\u6807\u7b7e\u7684\u6587\u7ae0\u3011\" target=\"_blank\" rel=\"noopener\">\u57fa\u7840\u5927\u6a21\u578b<\/a> Step 3 \u6b63\u5f0f<a href=\"https:\/\/www.1ai.net\/en\/tag\/%e5%bc%80%e6%ba%90\" title=\"\u3010\u67e5\u770b\u542b\u6709[\u5f00\u6e90]\u6807\u7b7e\u7684\u6587\u7ae0\u3011\" target=\"_blank\" rel=\"noopener\">\u5f00\u6e90<\/a>\uff0cStep 3 API \u5df2\u4e0a\u7ebf\u9636\u8dc3\u661f\u8fb0\u5f00\u653e\u5e73\u53f0\uff08platform.stepfun.com\uff09\uff0c\u7528\u6237\u4e5f\u53ef\u4ee5\u5728\u201c\u9636\u8dc3 AI\u201d\u5b98\u7f51\uff08stepfun.com\uff09\u548c\u201c\u9636\u8dc3 AI\u201dApp \u8fdb\u884c\u4f53\u9a8c\u3002<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"alignnone size-full wp-image-40565\" title=\"3e5f6b92j00t0ap720036d000u000gfp\" src=\"https:\/\/www.1ai.net\/wp-content\/uploads\/2025\/08\/3e5f6b92j00t0ap720036d000u000gfp.jpg\" alt=\"3e5f6b92j00t0ap720036d000u000gfp\" width=\"1080\" height=\"591\" \/><\/p>\n<p>\u636e\u4ecb\u7ecd\uff0cStep 3 \u7684\u591a\u6a21\u6001\u80fd\u529b\u56f4\u7ed5\u201c\u8f7b\u91cf\u89c6\u89c9\u8def\u5f84\u201d\u4e0e\u201c\u7a33\u5b9a\u534f\u540c\u8bad\u7ec3\u201d\u5c55\u5f00\uff0c\u91cd\u70b9\u89e3\u51b3\u89c6\u89c9\u5f15\u5165\u5e26\u6765\u7684 token \u8d1f\u62c5\u4e0e\u8bad\u7ec3\u5e72\u6270\u95ee\u9898\u3002\u4e3a\u6b64\uff0c\u5176\u91c7\u7528 5B Vision Encoder\uff0c\u5e76\u901a\u8fc7\u53cc\u5c42 2D \u5377\u79ef\u5bf9\u89c6\u89c9\u7279\u5f81\u8fdb\u884c\u964d\u91c7\u6837\uff0c\u5c06\u89c6\u89c9 token \u6570\u91cf\u51cf\u5c11\u5230\u539f\u6765\u7684 1\/16\uff0c\u51cf\u8f7b\u4e0a\u4e0b\u6587\u957f\u5ea6\u538b\u529b\uff0c\u63d0\u5347\u63a8\u7406\u6548\u7387\u3002<\/p>\n<p>1AI\u9644\u5b98\u65b9\u5bf9 Step 3 \u6a21\u578b\u7684\u4ecb\u7ecd\u5982\u4e0b\uff1a<\/p>\n<p>\u6838\u5fc3\u8981\u70b9<\/p>\n<ul>\n<li>Step 3 \u517c\u987e\u667a\u80fd\u4e0e\u6548\u7387\uff0c\u4e13\u4e3a\u8ffd\u6c42\u6027\u80fd\u4e0e\u6210\u672c\u6781\u81f4\u5747\u8861\u7684\u4f01\u4e1a\u548c\u5f00\u53d1\u8005\u8bbe\u8ba1\uff0c\u65e8\u5728\u9762\u5411\u63a8\u7406\u65f6\u4ee3\u6253\u9020\u6700\u9002\u5408\u5e94\u7528\u7684\u6a21\u578b\u3002<\/li>\n<li>Step 3 \u91c7\u7528 MoE \u67b6\u6784\uff0c\u603b\u53c2\u6570\u91cf 321B\uff0c\u6fc0\u6d3b\u53c2\u6570\u91cf 38B\u3002<\/li>\n<li>Step 3 \u62e5\u6709\u5f3a\u5927\u7684\u89c6\u89c9\u611f\u77e5\u548c\u590d\u6742\u63a8\u7406\u80fd\u529b\uff0c\u53ef\u51c6\u786e\u5b8c\u6210\u8de8\u9886\u57df\u7684\u590d\u6742\u77e5\u8bc6\u7406\u89e3\u3001\u6570\u5b66\u4e0e\u89c6\u89c9\u4fe1\u606f\u7684\u4ea4\u53c9\u5206\u6790\uff0c\u4ee5\u53ca\u65e5\u5e38\u751f\u6d3b\u4e2d\u7684\u5404\u7c7b\u89c6\u89c9\u5206\u6790\u95ee\u9898\u3002<\/li>\n<li>\u901a\u8fc7 MFA\uff08Multi-matrix Factorization Attention\uff09 &amp; AFD\uff08Attention-FFN Disaggregation\uff09\u7684\u4f18\u5316\uff0c\u5728\u5404\u7c7b\u82af\u7247\u4e0a\u63a8\u7406\u6548\u7387\u5747\u5927\u5e45\u63d0\u5347\u3002<\/li>\n<li>\u9762\u5411 AFD \u573a\u666f\u7684 StepMesh \u901a\u4fe1\u5e93\u5df2\u968f\u6a21\u578b\u4e00\u540c\u5f00\u6e90\uff0c\u63d0\u4f9b\u53ef\u8de8\u786c\u4ef6\u7684\u6807\u51c6\u90e8\u7f72\u63a5\u53e3\uff0c\u652f\u6301\u5173\u952e\u6027\u80fd\u5728\u5b9e\u9645\u670d\u52a1\u4e2d\u7684\u7a33\u5b9a\u590d\u73b0\u3002<\/li>\n<li>\u6a21\u578b\u9650\u65f6\u6298\u6263\u4e2d\uff0c\u6240\u6709\u8bf7\u6c42\u5747\u6309\u6700\u4f4e\u4ef7\u683c\u8ba1\u7b97\uff0c\u6bcf\u767e\u4e07 token \u4ef7\u683c\u4f4e\u81f3\u8f93\u5165 1.5 \u5143\uff0c\u8f93\u51fa 4 \u5143\u3002<\/li>\n<\/ul>\n<p>Step 3 API \u5df2\u4e0a\u7ebf\u9636\u8dc3\u661f\u8fb0\u5f00\u653e\u5e73\u53f0\uff08platform.stepfun.com\uff09\uff0c\u5927\u5bb6\u4e5f\u53ef\u4ee5\u5728\u201c\u9636\u8dc3 AI\u201d\u5b98\u7f51\uff08stepfun.com\uff09\u548c\u201c\u9636\u8dc3 AI\u201dApp\uff08\u5e94\u7528\u5546\u5e97\u641c\u7d22\u4e0b\u8f7d\uff09\u8fdb\u884c\u4f53\u9a8c\u3002<\/p>\n<p>\u884c\u4e1a\u9886\u5148\u7684\u6a21\u578b\u6027\u80fd<\/p>\n<p>\u6211\u4eec\u5728 MMMU\u3001MathVision\u3001SimpleVQA\u3001AIME 2025\u3001GPQA-Diamond\u3001LiveCodeBench \uff082024.08-2025.05\uff09\u7b49\u8bc4\u6d4b\u96c6\u4e0a\u5bf9 Step 3 \u8fdb\u884c\u4e86\u6d4b\u8bd5\uff0c\u5728\u540c\u7c7b\u578b\u5f00\u6e90\u6a21\u578b\u4e2d\uff0cStep 3 \u6210\u7ee9\u884c\u4e1a\u9886\u5148\u3002<\/p>\n<p>\u6280\u672f\u4eae\u70b9<\/p>\n<p>Step 3 \u91cd\u70b9\u89e3\u51b3\u591a\u6a21\u6001\u534f\u540c\u3001\u7cfb\u7edf\u89e3\u7801\u6210\u672c\u4e0e\u63a8\u7406\u6548\u7387\u95ee\u9898\uff0c\u5728\u8bad\u7ec3\u8def\u5f84\u3001\u67b6\u6784\u8bbe\u8ba1\u548c\u63a8\u7406\u90e8\u7f72\u4e0a\u505a\u4e86\u7cfb\u7edf\u7ea7\u4f18\u5316\u3002<\/p>\n<p><strong>1\u3001\u9884\u8bad\u7ec3\u67b6\u6784<\/strong><\/p>\n<p>Step 3 \u7684\u6838\u5fc3\u7ed3\u6784\u91c7\u7528\u81ea\u7814 MFA \u6ce8\u610f\u529b\u673a\u5236\uff0c\u6709\u6548\u964d\u4f4e\u6ce8\u610f\u529b\u8ba1\u7b97\u4e2d\u7684 KV \u7f13\u5b58\u5f00\u9500\u4e0e\u7b97\u529b\u6d88\u8017\u3002\u5728\u4e0d\u727a\u7272\u6a21\u578b\u80fd\u529b\u7684\u524d\u63d0\u4e0b\uff0c\u8fd9\u4e00\u65b9\u6848\u5b9e\u73b0\u4e86\u8d44\u6e90\u5229\u7528\u4e0e\u63a8\u7406\u6548\u7387\u7684\u5e73\u8861\uff0c\u4f7f\u5f97\u6a21\u578b\u53ef\u5728 8\u00d748GB \u663e\u5361\u4e0a\u5b8c\u6210\u5927\u541e\u5410\u91cf\u63a8\u7406\uff0c\u5177\u5907\u771f\u5b9e\u90e8\u7f72\u7684\u53ef\u884c\u6027\u3002<\/p>\n<p><strong>2\u3001\u591a\u6a21\u6001\u9884\u8bad\u7ec3<\/strong><\/p>\n<p>Step 3 \u7684\u591a\u6a21\u6001\u80fd\u529b\u56f4\u7ed5\u201c\u8f7b\u91cf\u89c6\u89c9\u8def\u5f84\u201d\u4e0e\u201c\u7a33\u5b9a\u534f\u540c\u8bad\u7ec3\u201d\u5c55\u5f00\uff0c\u91cd\u70b9\u89e3\u51b3\u89c6\u89c9\u5f15\u5165\u5e26\u6765\u7684 token \u8d1f\u62c5\u4e0e\u8bad\u7ec3\u5e72\u6270\u95ee\u9898\u3002\u4e3a\u6b64\uff0c\u6211\u4eec\u91c7\u7528 5B Vision Encoder\uff0c\u5e76\u901a\u8fc7\u53cc\u5c42 2D \u5377\u79ef\u5bf9\u89c6\u89c9\u7279\u5f81\u8fdb\u884c\u964d\u91c7\u6837\uff0c\u5c06\u89c6\u89c9 token \u6570\u91cf\u51cf\u5c11\u5230\u539f\u6765\u7684 1\/16\uff0c\u51cf\u8f7b\u4e0a\u4e0b\u6587\u957f\u5ea6\u538b\u529b\uff0c\u63d0\u5347\u63a8\u7406\u6548\u7387\u3002<\/p>\n<p>\u4e3a\u4fdd\u8bc1\u591a\u6a21\u6001\u8bad\u7ec3\u7684\u7a33\u5b9a\u6027\uff0c\u8bad\u7ec3\u8fc7\u7a0b\u5206\u4e3a\u4e24\u4e2a\u9636\u6bb5\uff1a\u7b2c\u4e00\u9636\u6bb5\u5f3a\u5316 Encoder \u611f\u77e5\uff0c\u7b2c\u4e8c\u9636\u6bb5\u51bb\u7ed3\u89c6\u89c9\u7f16\u7801\u5668\uff0c\u4ec5\u4f18\u5316\u4e3b\u5e72\u4e0e\u8fde\u63a5\u5c42\uff0c\u4ee5\u51cf\u5c11\u68af\u5ea6\u5e72\u6270\u3002\u8bad\u7ec3\u8bed\u6599\u4e5f\u9700\u4e0e\u7b56\u7565\u5339\u914d\uff0c\u4fdd\u969c\u7a33\u5b9a\u534f\u540c\u3002\u591a\u6a21\u8bed\u6599\u6db5\u76d6 Pair\u3001Interleave \u4e0e\u591a\u4efb\u52a1\u6570\u636e\uff0c\u5728\u6e05\u6d17\u73af\u8282\u4e2d\u5f15\u5165\u76f8\u4f3c\u5ea6\u8fc7\u6ee4\u3001\u91cd\u91c7\u6837\u4e0e\u4efb\u52a1\u6bd4\u4f8b\u63a7\u5236\uff0c\u8fdb\u4e00\u6b65\u63d0\u5347\u56fe\u6587\u534f\u540c\u8d28\u91cf\u4e0e\u8bad\u7ec3\u9c81\u68d2\u6027\u3002<\/p>\n<p><strong>3\u3001AFD \u89e3\u8026\u7cfb\u7edf<\/strong><\/p>\n<p>Step 3 \u5728\u7cfb\u7edf\u67b6\u6784\u5c42\u91cd\u6784\u4e86\u89e3\u7801\u6d41\u7a0b\uff0c\u91cd\u70b9\u89e3\u51b3 Attention \u4e0e FFN \u6df7\u5408\u6267\u884c\u5e26\u6765\u7684\u63a8\u7406\u74f6\u9888\u4ee5\u53ca\u8d44\u6e90\u4e0d\u5339\u914d\u95ee\u9898\u3002\u4e3a\u6b64\uff0c\u6211\u4eec\u5b9e\u73b0\u4e86\u9ad8\u6027\u80fd\u7684 AFD\uff08Attention-FFN Disaggregation\uff09\u65b9\u6848\uff0c\u5c06\u4e24\u7c7b\u8ba1\u7b97\u4efb\u52a1\u89e3\u8026\u6210\u4e3a\u4e24\u4e2a\u5b50\u7cfb\u7edf\uff0c\u5e76\u901a\u8fc7\u591a\u7ea7\u6d41\u6c34\u7ebf\u5e76\u884c\u8c03\u5ea6\uff0c\u6709\u6548\u63d0\u5347\u6574\u4f53\u541e\u5410\u6548\u7387\u3002<\/p>\n<p>\u7531\u4e8e\u89e3\u8026\u540e\u7684\u5b50\u7cfb\u7edf\u4e4b\u95f4\u5bf9\u6570\u636e\u4f20\u8f93\u6709\u6781\u9ad8\u8981\u6c42\uff0c\u6211\u4eec\u540c\u65f6\u7814\u53d1\u4e86\u9762\u5411 AFD \u573a\u666f\u7684 StepMesh \u901a\u4fe1\u5e93\uff0c\u57fa\u4e8e GPU Direct RDMA \u5b9e\u73b0\u8de8\u5361\u7684\u4f4e\u5ef6\u8fdf\u548c\u9ad8\u5e26\u5bbd\u4f20\u8f93\uff0c\u540c\u65f6\u517c\u5907\u4e0d\u5360\u7528 GPU \u8ba1\u7b97\u8d44\u6e90\u3001\u9002\u914d\u591a\u7c7b\u5f02\u6784\u786c\u4ef6\u7b49\u4f18\u52bf\u3002\u5728 50ms \u89e3\u7801\u7684 SLA \u524d\u63d0\u4e0b\uff0cStep 3 \u5728 Hopper GPU \u4e0a\u7684\u541e\u5410\u8fbe\u5230 4039 token \/ gpu \/ s\uff0c\u663e\u8457\u9ad8\u4e8e\u7c7b\u4f3c\u8bbe\u7f6e\u4e0b\u7684 DeepSeek V3\uff082324 token \/ gpu \/ s\uff09\uff0c\u4e14\u8be5\u6027\u80fd\u589e\u76ca\u5728\u7279\u5b9a\u786c\u4ef6\u4e0e\u957f\u6587\u573a\u666f\u4f1a\u8fdb\u4e00\u6b65\u653e\u5927\u81f3 300%\u3002<\/p>\n<p>StepMesh \u5e93\u5df2\u968f\u6a21\u578b\u4e00\u540c\u5f00\u6e90\uff0c\u63d0\u4f9b\u53ef\u8de8\u786c\u4ef6\u7684\u6807\u51c6\u90e8\u7f72\u63a5\u53e3\uff0c\u652f\u6301\u5173\u952e\u6027\u80fd\u5728\u5b9e\u9645\u670d\u52a1\u4e2d\u7684\u7a33\u5b9a\u590d\u73b0\u3002\u4e4b\u540e\u6211\u4eec\u4e5f\u4f1a\u79ef\u6781\u4e0e\u5404\u5f00\u6e90\u793e\u533a\u5408\u4f5c\u63a8\u5e7f\uff0c\u8ba9\u4ee5\u4e0a\u6280\u672f\u66f4\u5bb9\u6613\u88ab\u91c7\u7eb3\u548c\u4f7f\u7528\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>August 1, 2012 - Stepfun announced that Step 3, a new generation of basic big model, is officially open source, and the Step 3 API has been launched on Stepfun's open platform (platform.stepfun.com), and users can also experience it on the official website of Stepfun AI (stepfun.com) and the Stepfun App. \"Users can also experience Step 3 on the Step AI website (stepfun.com) and the Step AI App. According to the introduction, the multimodal capability of Step 3 is centered on \"lightweight visual path\" and \"stable collaborative training\", focusing on solving the token burden and training interference caused by the introduction of vision. To this end, it adopts 5B Vision Encoder and downsamples visual features through two-layer 2D convolution to reduce the number of visual tokens to the original<\/p>","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[146],"tags":[6731,219,1893],"collection":[],"class_list":["post-40564","post","type-post","status-publish","format-standard","hentry","category-news","tag-6731","tag-219","tag-1893"],"acf":[],"_links":{"self":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/posts\/40564","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/comments?post=40564"}],"version-history":[{"count":0,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/posts\/40564\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/media?parent=40564"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/categories?post=40564"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/tags?post=40564"},{"taxonomy":"collection","embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/collection?post=40564"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}