[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fjUyODeY9poNdBuDkNMFKsyx0gGO-0NoCRRyWs7Y5vyI":3},{"code":4,"msg":5,"data":6},200,"操作成功",{"id":7,"title":8,"content":9,"digest":10,"source":10,"coverPath":11,"thumbsCoverPath":12,"isTop":13,"isShow":14,"baseClick":13,"clickCount":15,"createTime":16,"typeId":17,"isNewest":18,"newsInfoTypeRespVo":19,"voiceUrl":22,"voiceSize":23,"taskId":24,"releaseTime":25,"titleEn":26,"contentEn":27,"voiceUrlEn":28,"taskIdEn":29,"voiceSizeEn":30},1529," 小米MoE大模型MiMo-V2-Flash发布，性能超越顶尖竞争者！","\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px; color: rgb(255, 153, 0);\">小米今日正式发布了最新一代的MoE架构大模型——MiMo-V2-Flash。这款模型总参数规模达3090亿，激活参数150亿，展现出在推理、编码及智能体应用场景中的显著优势。作为小米在人工智能领域的重要布局，MiMo-V2-Flash的发布不仅是技术的突破，也是市场竞争的加剧。\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">MiMo-V2-Flash支持混合思维模式，用户可以根据需求自由切换深度思考与即时响应模式。这种灵活性使得模型在实际应用中能够更好地满足用户需求，提升工作效率。此外，该模型还具备一键生成完整HTML网页的能力，并能够与ClaudeCode、Cursor等主流编码框架无缝协作，进一步增强了其在软件开发领域的应用潜力。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">在性能测试中，MiMo-V2-Flash与DeepSeek-V3.2形成直接竞争关系。根据基准数据显示，该模型在数学竞赛AIME2025和科学知识GPQA-Diamond测试中均位列开源模型前二。在软件工程能力SWE验证及多语言测试中，更是超越了所有开源竞品，性能表现与全球顶尖闭源模型相持平。这一系列数据无疑为小米在AI领域的实力提供了有力证明。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">特别值得一提的是，MiMo-V2-Flash在长上下文处理方面的表现已超越K2Thinking模型。在SWE-BenchVerified测试中，其解决率达71.7%，而在BrowseComp搜索评估中得分45.4，经过上下文管理后提升至58.3。这些成绩显示出MiMo-V2-Flash在处理复杂信息时的强大能力，适合更多高需求场景的应用。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">在技术架构上，MiMo-V2-Flash采用了全局注意力（GA）与滑动窗口注意力（SWA）的1:5混合结构，这一创新设计既保持了线性注意力的计算效率，又显著提升了长文本处理能力。此外，模型引入的多词元预测（MTP）训练技术，通过同步生成多个候选token并并行验证的方式，使解码吞吐量提升了2-2.6倍。这一技术的应用，无疑为模型的整体性能提供了支持。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">在后训练阶段，MiMo-V2-Flash采用了多教师在线策略蒸馏（MOPD）方法，计算资源的需求仅为传统方法的1\u002F50，却能达到同等的优化效果。这种高效的训练方式形成了“教学-学习”的闭环迭代机制，为模型的持续优化奠定了基础。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">实际场景测试中，MiMo-V2-Flash展现了其多样化的能力。在网页开发任务中，它能够生成包含商品轮播、规格选择等功能的电商页面，同时也支持手势交互的3D应用；在创意生成方面，模型能够创作忧郁爱情故事和非虚构社会观察作品。此外，在智能体交互测试中，MiMo-V2-Flash不仅可以解答哲学问题，还能编写科幻悬疑剧本，显示出其在多领域应用的潜力。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">然而，实测中也发现，部分复杂交互场景（如教育类太阳系探索器）存在稳定性问题，需要多次生成才能达到预期效果。尽管如此，MiMo-V2-Flash的全面开源及推理代码的贡献至开发者社区SGLang，标志着小米在AI领域的开放态度与技术共享精神。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">在API服务方面，小米为用户开启了限时免费体验，吸引更多开发者参与到这一技术生态中。根据技术文档，在Prefill单机吞吐50000toks\u002Fs条件下，16K上下文长度的解码吞吐量可达5000-15000toks\u002Fs，单请求吞吐量达到151-115toks\u002Fs。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">在价格策略方面，MiMo-V2-Flash的输入token单价为0.7元\u002F百万，输出token单价为2.1元\u002F百万，显著低于行业平均水平。这一价格策略不仅有助于提升用户的使用体验，也进一步推动了小米AI业务的市场竞争力。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">小米在AI领域的持续投入和突破，标志着其研发进入快车道。近期，小米集团宣布将AI与现实世界深度结合列为未来十年核心战略，AI业务投入连续四个季度环比增长超50%。为加速技术突破，小米同步启动了全球人才招募计划，设立千万元级岗位薪酬，力求引进大模型领域的顶尖人才。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">在过去的三个月内，小米技术团队已密集发布多篇学术论文，并陆续开源多个预训练模型，构建起完整的技术生态体系。随着小米在AI领域的不断发力，未来的科技趋势无疑值得我们期待。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"color: rgb(136, 136, 136);\">【新闻来源】搜狐新闻 \u003C\u002Fspan>\u003Ca href=\"https:\u002F\u002Fwww.sohu.com\u002Fa\u002F966451006_121956424\" rel=\"noopener noreferrer\" target=\"_blank\" style=\"color: rgb(136, 136, 136);\"> https:\u002F\u002Fwww.msn.cn\u002Fzh-cn\u002Fnews\u002Fother\u002Far-AA1RTDQy?ocid=BingHp01&cvid=6936317f054647a2afcd53fafcde084a&ei\u003C\u002Fa>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"color: rgb(136, 136, 136);\">（本网转发此文章，旨在为读者提供更多的信息资讯，所涉内容不构成投资、消费建议。文章事实如有疑问，请与有关方核实，文章观点非本网观点，仅供读者参考。）\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>","","https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F12\u002F289b657e864d49a986394dd2842d3385\u002FAI领域.jpg","https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F12\u002Fthumbs\u002F289b657e864d49a986394dd2842d3385\u002FAI领域.jpg",0,1,43,"2025-12-22 08:45",2,false,{"id":17,"name":20,"enName":21},"芯位视野","Xinwei Vision","https:\u002F\u002Fxinwei-dev-test.oss-cn-shenzhen.aliyuncs.com\u002Fintelligent\u002Faudio%3A6e9949d4-f831-43be-a972-ea7d405217da%3A0.wav?Expires=1766372167&OSSAccessKeyId=LTAI5tNvY2RkKjZw4LLWsrPK&Signature=IX%2BPYiz4ilOybIDi0tfu86jx4gs%3D",8763108,"6e9949d4-f831-43be-a972-ea7d405217da","2025-12-22 08:43"," Xiaomi MoE large model MiMo-V2-Flash is released, with performance surpassing top competitors!","\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px; color: rgb(255, 153, 0);\">Xiaomi officially released its latest generation MoE architecture large model - MiMo-V2-Flash today. This model has a total parameter scale of 309 billion, with 15 billion activated parameters, showing significant advantages in reasoning, coding, and intelligent agent application scenarios. As an important layout in the field of artificial intelligence, the release of MiMo-V2-Flash is not only a technological breakthrough but also a further intensification of market competition.\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">MiMo-V2-Flash supports a hybrid thinking mode, allowing users to freely switch between deep thinking and instant response modes according to their needs. This flexibility enables the model to better meet user needs in practical applications, improving work efficiency. In addition, this model has the ability to generate complete HTML web pages at one click and can seamlessly collaborate with mainstream coding frameworks such as ClaudeCode and Cursor, further enhancing its application potential in the software development field.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">In performance tests, MiMo-V2-Flash directly competes with DeepSeek-V3.2. According to benchmark data, this model ranks in the top two among open-source models in the AIME2025 math competition and the GPQA-Diamond science knowledge test. In the software engineering capability SWE verification and multi-language tests, it even surpasses all open-source competitors, with performance comparable to the world's top closed-source models. These data undoubtedly provide strong evidence for Xiaomi's strength in the AI field.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">It is particularly worth mentioning that MiMo-V2-Flash's performance in long context processing has surpassed the K2Thinking model. In the SWE-BenchVerified test, its solution rate reached 71.7%, and in the BrowseComp search evaluation, it scored 45.4, which was improved to 58.3 after context management. These achievements demonstrate the powerful capability of MiMo-V2-Flash in handling complex information, making it suitable for more high-demand application scenarios.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">In terms of technical architecture, MiMo-V2-Flash adopts a 1:5 mixed structure of global attention (GA) and sliding window attention (SWA), an innovative design that maintains the computational efficiency of linear attention while significantly improving the ability to process long texts. In addition, the model introduces the multi-token prediction (MTP) training technology, which enhances decoding throughput by 2-2.6 times by simultaneously generating multiple candidate tokens and verifying them in parallel. The application of this technology undoubtedly supports the overall performance of the model.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">In the post-training phase, MiMo-V2-Flash adopts the multi-teacher online strategy distillation (MOPD) method, which requires only 1\u002F50 of the computing resources of traditional methods but achieves the same level of optimization. This efficient training method forms a \"teaching-learning\" closed-loop iteration mechanism, laying the foundation for continuous model optimization.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">In practical scenario testing, MiMo-V2-Flash demonstrated its diverse capabilities. In web development tasks, it can generate e-commerce pages with functions such as product carousels and specification selections, and also supports 3D applications with gesture interaction. In creative generation, the model can create melancholic love stories and non-fiction social observation works. Additionally, in intelligent agent interaction tests, MiMo-V2-Flash can not only answer philosophical questions but also write science fiction mystery scripts, showing its potential for multi-field applications.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">However, during actual testing, some complex interactive scenarios (such as educational solar system explorers) had stability issues, requiring multiple generations to achieve the desired effect. Despite this, the comprehensive open source of MiMo-V2-Flash and the contribution of its inference code to the developer community SGLang mark Xiaomi's open attitude and spirit of technology sharing in the AI field.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">In terms of API services, Xiaomi has opened up limited-time free experience for users, attracting more developers to participate in this technological ecosystem. According to the technical documentation, under the condition of Prefill single-machine throughput of 50,000 toks\u002Fs, the decoding throughput for a 16K context length can reach 5,000-15,000 toks\u002Fs, and the throughput per request reaches 151-115 toks\u002Fs.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">In terms of pricing strategy, the input token unit price of MiMo-V2-Flash is 0.7 yuan per million, and the output token unit price is 2.1 yuan per million, significantly lower than the industry average. This pricing strategy not only helps improve user experience but also further enhances Xiaomi's market competitiveness in AI business.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Xiaomi's continuous investment and breakthroughs in the AI field mark its R&D entering a fast lane. Recently, Xiaomi Group announced that integrating AI with the real world will be the core strategy for the next ten years, and AI business investment has increased by more than 50% quarter-on-quarter for four consecutive quarters. To accelerate technological breakthroughs, Xiaomi has also launched a global talent recruitment plan, setting up positions with a salary of millions of yuan, aiming to attract top talents in the field of large models.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">In the past three months, Xiaomi's technical team has released multiple academic papers and successively open-sourced several pre-trained models, building a complete technical ecosystem. With Xiaomi's continued efforts in the AI field, the future of technological trends is undoubtedly worth looking forward to.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"color: rgb(136, 136, 136);\"> [News Source] Sohu News \u003C\u002Fspan>\u003Ca href=\"https:\u002F\u002Fwww.sohu.com\u002Fa\u002F966451006_121956424\" rel=\"noopener noreferrer\" target=\"_blank\" style=\"color: rgb(136, 136, 136);\"> https:\u002F\u002Fwww.msn.cn\u002Fzh-cn\u002Fnews\u002Fother\u002Far-AA1RTDQy?ocid=BingHp01&cvid=6936317f054647a2afcd53fafcde084a&ei\u003C\u002Fa>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"color: rgb(136, 136, 136);\"> (This article is forwarded by this site to provide readers with more information and news. The content involved does not constitute investment or consumption advice. If there are any doubts about the facts of the article, please verify with the relevant parties. The views of the article are not the views of this site, and are for reference only.)\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>","https:\u002F\u002Fxinwei-dev-test.oss-cn-shenzhen.aliyuncs.com\u002Fintelligent\u002Faudio%3A5e833b87-6f6e-4856-b3eb-66c28d53931b%3A0.wav?Expires=1774838439&OSSAccessKeyId=LTAI5tNvY2RkKjZw4LLWsrPK&Signature=tDRPAbejFfTeEpVowAusURVN5Rs%3D","5e833b87-6f6e-4856-b3eb-66c28d53931b",11397402]