[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fOos0rZLX54FqU5oWYvAsJNtWyaF5_Han-ImaQntIT0w":3,"$fWNsfe5LyMk_QM7fXcoufrWVu1aa3Gn4kGdkU0eJxsuI":89},{"id":4,"name":5,"code":6,"version":7,"logo":8,"url":9,"vendorId":10,"vendorName":11,"vendorCode":12,"vendorCountry":13,"modelType":35,"openSource":18,"openLicense":36,"paramSize":37,"contextLength":38,"intro":39,"supportApi":18,"apiUrl":40,"pricingType":35,"freeQuota":41,"score":42,"isRecommend":43,"isHot":18,"click":44,"tags":45,"content":46,"pricings":47,"scores":61},52,"MiniMax-M3","minimax-m3","M3","\u002Flogos\u002Fminimax.svg","https:\u002F\u002Fchat.minimaxi.com",32,"MiniMax 稀宇科技","minimax",[14,19,23,27,31],{"id":15,"name":16,"color":17,"type":18},4,"开源","#909399",1,{"id":20,"name":21,"color":22,"type":18},5,"支持API","#00B4A0",{"id":24,"name":25,"color":26,"type":18},8,"国产自研","#F56C6C",{"id":28,"name":29,"color":30,"type":18},10,"长上下文","#3B55D9",{"id":32,"name":33,"color":34,"type":18},11,"多模态理解","#9C27B0",2,"Apache-2.0","428B(A23B)",1000000,"MiniMax 原生多模态开源旗舰，首个从 Step 0 做多模态混合训练的开源模型","https:\u002F\u002Fplatform.minimaxi.com","详见官网当期免费额度活动",8.8,0,18,[],"MiniMax M3 于 2026 年 6 月发布并于 6 月 12 日开源模型权重，同步发布 MSA（MiniMax Sparse Attention）技术论文。总参数 428B、激活参数 23B（MoE 架构），最高支持 100 万 token 上下文，512K 以内为稳定可用区间，单次最大输出 64K。它最大的不一样在于训练方式——是第一个从 Step 0 就开始做多模态混合训练的开源模型：预训练阶段就把文本、图像与其他模态数据交错排列，让不同模态的语义空间从底层就完成融合，而不是在纯文本模型上后挂一个视觉模块。发布两周后即在 Artificial Analysis 综合智能指数排行榜上取得全球开源模型的最高排名，输出速度从 30 TPS 提升至约 80 TPS。生态适配是其另一个亮点：与 OpenClaw、Hermes Agent、Multica、DeerFlow 等项目深度合作，并与华为昇腾、摩尔线程、沐曦、昆仑芯、海光、NVIDIA、AMD 等芯片平台实现 Day 0 适配。后续推出的 M3.1-Flash 进一步把价格压到输入 1 元 \u002F 输出 4 元，本文一并归入本条目。",[48,56],{"itemName":49,"billingType":18,"inputPrice":50,"outputPrice":51,"price":43,"unit":52,"currency":53,"freeQuota":54,"intro":55},">512K 上下文",4.2,16.8,"元\u002F百万tokens","CNY",null,"缓存读取 0.84 元",{"itemName":57,"billingType":18,"inputPrice":58,"outputPrice":59,"price":43,"unit":52,"currency":53,"freeQuota":54,"intro":60},"≤512K 上下文",2.1,8.4,"缓存读取 0.42 元",[62,66,69,73,77,80,84],{"metricId":18,"metricName":63,"metricCode":64,"maxScore":28,"score":65},"中文理解","chinese",8.5,{"metricId":35,"metricName":67,"metricCode":68,"maxScore":28,"score":24},"文本写作","writing",{"metricId":70,"metricName":71,"metricCode":72,"maxScore":28,"score":65},3,"逻辑推理","reasoning",{"metricId":15,"metricName":74,"metricCode":75,"maxScore":28,"score":76},"代码能力","coding",9,{"metricId":20,"metricName":78,"metricCode":79,"maxScore":28,"score":76},"多模态","multimodal",{"metricId":81,"metricName":82,"metricCode":83,"maxScore":28,"score":76},6,"长文本","long-context",{"metricId":85,"metricName":86,"metricCode":87,"maxScore":28,"score":88},7,"性价比","cost-performance",9.5,[90,92,94,98,99,100,103,106],{"id":20,"name":21,"color":22,"type":18,"toolCount":91},35,{"id":28,"name":29,"color":30,"type":18,"toolCount":93},33,{"id":35,"name":95,"color":96,"type":18,"toolCount":97},"免费增值","#409EFF",21,{"id":15,"name":16,"color":17,"type":18,"toolCount":44},{"id":32,"name":33,"color":34,"type":18,"toolCount":28},{"id":76,"name":101,"color":102,"type":18,"toolCount":28},"推理强化","#E65100",{"id":70,"name":104,"color":105,"type":18,"toolCount":85},"付费使用","#E6A23C",{"id":24,"name":25,"color":26,"type":18,"toolCount":85}]