广告
加载中

小米发布MiMo-V2.6大模型 登顶开源AI性能榜单

亿邦AI 2026-09-23 11:38
亿邦AI 2026/09/23 11:38

邦小白快读

EN
全文速览

MiMo-V2.6登顶开源AI性能榜,模型可免费下载使用。

1. 高性能免费获取:Pro和Flash模型权重已在Hugging Face开放,MIT许可允许免费下载、定制和商用;开发者平台OpenCode已上线,Flash版本对用户免费开放一周。

2. 低价格API:Pro定价每百万输入0.435美元、输出0.87美元;Flash更便宜,缓存输入仅0.0028美元,适合高并发场景。

3. 实用表现:支持文本、图像、音频、视频多模态输入和百万级上下文,可完成桌面软件操作、前端设计、音视频制作等长流程任务。

4. 风险提醒:Anthropic指控小米曾绕过限制蒸馏Claude模型能力并用于训练,相关争议尚未有官方回应,普通用户在使用时需留意合规风险。

小米通过大模型技术登顶、开源低价策略塑造科技品牌形象,并强化产品矩阵布局。

1. 品牌营销:以智能指数第一、同比闭源模型低成本的优势制造传播话题,提升品牌科技感与行业话语权。

2. 产品研发与定位:推出Pro、Flash、UltraSpeed多个版本,覆盖高智能、高并发、超高速场景,反映AI产品差异化研发思路。

3. 定价与价格竞争:API定价显著低于同性能闭源模型,配合开源权重免费下载,形成以价格和开放性抢生态的竞争策略。

4. 消费趋势观察:开发者可低成本体验并商用模型,可能带动更多企业级AI应用需求;同时“非法蒸馏”指控带来的品牌合规风险也值得注意。

MiMo-V2.6为卖家提供低成本AI接入机会,也伴随合规风险与商业模式参考。

1. 机会提示:模型采用MIT许可免费开源,API定价极低,商家可用于客服、内容生成、营销文案等场景,降低AI应用成本。

2. 商业应用:支持多模态和百万token上下文,可协调智能体完成桌面软件操作、前端设计、音视频制作,适合开发自动化工作流。

3. 合作方式:可直接从Hugging Face下载权重部署,或通过API接入;OpenCode平台已上线,Flash版本限时免费,便于试用。

4. 风险提示:Anthropic将小米列入滥用Claude模型名单,涉及“非法蒸馏”指控,卖家选用相关模型服务时需评估供应链合规与法律风险。

5. 商业模式参考:小米以“免费权重+低价API”双轨运行、用开源生态反哺品牌的技术商业策略值得学习。

MiMo-V2.6开源低门槛,为工厂生产自动化与数字化提供可用AI方案。

1. 生产应用:演示中支持机械臂视觉控制,可用于工厂自动化操作;长流程智能体能完成桌面软件操作等任务,辅助生产流程管理。

2. 数据与部署:模型权重可免费下载,MIT许可允许部署至自有或租赁硬件,工厂可保护生产数据不外传,按需定制模型。

3. 设计需求:模型支持多模态输入和长上下文,可完成Blender场景搭建、3D世界生成、前端设计等,辅助产品设计与数字孪生搭建。

4. 成本启示:训练仅需数天、成本数百万美元级,且使用国产芯片,整体造芯与训练成本可控,显示AI基础设施成本在下降,有利于制造企业跟进数字化改造。

MiMo-V2.6以高性能、低成本、强开放性和训练框架重构服务商的技术底座。

1. 行业趋势:开源模型综合得分追平顶级闭源模型,证明开源闭源差距缩小,未来更多客户会考虑开源方案;强化学习成为性能跃升关键。

2. 新技术亮点:MoE稀疏架构、百万上下文、多模态输入;训练使用You Only RL Once策略与异步组相对策略优化,并公开技术报告和框架。

3. 客户痛点解决:客户常困于AI成本高、数据隐私和定制难。MiMo开放权重+MIT许可支持本地部署,API定价远低于闭源模型,Flash版适合高并发低成本场景,UltraSpeed版提供20倍高速输出。

4. 解决方案扩展:提供7000余个带自动评分器的训练任务和90亿参数小模型,服务商可基于此快速搭建针对软件、网络安全、办公等行业的工具链。

MiMo-V2.6系列的低价高性能为平台引入开发者流量和合规风险带来双重课题。

1. 平台集成机会:开发者平台OpenCode已上线该系列模型,Flash版免费一周吸引试用;平台可效仿通过大模型上线扩大开发者生态。

2. 运营与成本:Flash版API定价为全球第二低前沿模型,缓存输入超低价,适合平台作为默认模型或长上下文任务处理;多模态长上下文特性也可支持平台复杂应用。

3. 风险规避:Anthropic报告指控小米等中国AI实验室“非法蒸馏”Claude能力,平台接入小米模型前需建立合规审查与风控机制,关注后续回应及监管动态。

4. 招商与管理参考:小米以开源权重+免费试用+社区平台方式推广模型,平台可借鉴此模式联合模型厂商运营开发者活动。

MiMo-V2.6提供了稀缺的规模化强化学习训练技术细节和开源资源,并引发模型版权与合规讨论。

1. 产业新动向:开源模型首次在智能指数总榜登顶并追平顶级闭源,显示后训练强化学习成为能力提升的主要驱动力;完全基于国产芯片实现低成本大规模训练值得关注。

2. 技术方法:训练采用You Only RL Once策略,将多领域任务与轻量智能体脚手架整合单次训练,配合异步组相对策略优化提升GPU利用率;专门设计“作弊检测智能体”、冻结路由机制保障奖励可靠性,作弊轨迹占比低于2%。

3. 开放资源:论文级技术报告、端到端训练框架、7000余个自动评分器任务、90亿参数蒸馏小模型均已开放,利于复现和研究。

4. 争议与法律问题:Anthropic指控小米通过OpenClaw、OpenCode工具将用户对话汇入Claude产生40万轮交互,涉嫌“非法蒸馏”;该事件揭示大模型训练数据来源、服务条款合规及跨境AI治理问题,需要政策法规层面关注。

返回默认

声明:快读内容全程由AI生成,请注意甄别信息。如您发现问题,请发送邮件至 run@ebrun.com 。

我是 品牌商 卖家 工厂 服务商 平台商 研究者 帮我再读一遍。

Quick Summary

MiMo-V2.6 tops the open-source AI performance leaderboard, and the model is free to download and use.

1. High performance, free access: The Pro and Flash model weights are now open on Hugging Face under an MIT license, allowing free download, customization, and commercial use. The OpenCode developer platform has also launched, with the Flash version free to users for one week.

2. Low-cost API: Pro is priced at $0.435 per million input tokens and $0.87 per million output tokens; Flash is even cheaper, with cached input at just $0.0028 per million tokens, making it well suited to high-concurrency scenarios.

3. Practical performance: The model supports multimodal input covering text, images, audio, and video, with a million-level context window. It can handle long-horizon tasks such as desktop software operation, front-end design, and audio/video production.

4. Risk note: Anthropic has accused Xiaomi of circumventing restrictions to distill Claude model capabilities for training. There has been no official response to the dispute, and general users should remain mindful of compliance risks.

Xiaomi is leveraging the MiMo-V2.6 top ranking, open-source, and low-price strategy to strengthen its technology brand image and expand its product matrix.

1. Brand marketing: Topping the intelligence index and offering lower costs than comparable closed-source models generates buzz, reinforcing Xiaomi's technological credibility and industry influence.

2. Product development and positioning: The launch of Pro, Flash, and UltraSpeed versions covers high-intelligence, high-concurrency, and ultra-fast scenarios, reflecting a differentiated approach to AI product R&D.

3. Pricing and competition: API pricing is significantly lower than closed-source models of similar performance, and free open-source weights further support a go-to-market strategy built on price and openness to capture the ecosystem.

4. Consumer trend watch: Developers can now experience and commercialize the model at low cost, potentially fueling enterprise AI adoption; meanwhile, the "illegal distillation" allegation introduces brand and compliance risk that bears monitoring.

MiMo-V2.6 gives sellers a low-cost entry point to AI adoption, but also brings compliance risk and a useful business model reference.

1. Opportunity: The model is fully open source under an MIT license with extremely low API pricing. Merchants can apply it to customer service, content generation, and marketing copy, lowering the cost of AI adoption.

2. Commercial applications: With multimodal input and a million-token context window, agents can orchestrate desktop software operations, front-end design, and audio/video production, making it well suited to automated workflows.

3. Access options: Weights can be downloaded directly from Hugging Face for self-deployment, or the model can be accessed via API. The OpenCode platform is live, and the Flash version is free for a limited time, making trial easy.

4. Risk warning: Anthropic has placed Xiaomi on a list of alleged Claude abusers over "illegal distillation" claims. Sellers adopting Xiaomi's models should assess supply-chain compliance and legal exposure.

5. Business model reference: Xiaomi's dual-track strategy of "free weights + low-cost API" and using its open ecosystem to reinforce the brand's technology positioning is a model worth studying.

MiMo-V2.6's open-source, low-barrier design offers factories a practical AI option for production automation and digitalization.

1. Production applications: The model has been demonstrated controlling robotic arms through visual input, supporting factory automation; its long-horizon agentic capabilities can also handle desktop software tasks, assisting with production workflow management.

2. Data and deployment: Model weights are free to download, and the MIT license permits deployment on owned or rented hardware. Factories can keep production data in-house while customizing the model as needed.

3. Design capabilities: The model supports multimodal input and long context, and can complete Blender scene construction, 3D world generation, and front-end design, aiding product design and digital twin development.

4. Cost implications: Training took only days and cost on the order of millions of dollars, using domestic chips. The manageable total cost of chip and model development shows AI infrastructure costs are declining, which supports manufacturers pursuing digital upgrades.

MiMo-V2.6 reshapes the service provider technology stack with high performance, low cost, strong openness, and a new training framework.

1. Industry trend: An open-source model has matched top-tier closed-source models on an aggregate benchmark score, confirming that the gap is narrowing and that more customers will consider open-source options. Reinforcement learning is emerging as the key driver of performance gains.

2. Technical highlights: The model uses a sparse MoE architecture with a million-token context and multimodal input. Training employed the You Only RL Once strategy combined with asynchronous group-relative policy optimization, with the technical report and framework published openly.

3. Solving client pain points: Clients are often constrained by high AI costs, data privacy concerns, and limited customization. MiMo's open weights and MIT license support local deployment, while API pricing sits far below that of closed-source models. The Flash tier suits high-concurrency, low-cost workloads, and UltraSpeed delivers 20x faster output.

4. Expanding solution offerings: With more than 7,000 training tasks equipped with automatic graders and a 9-billion-parameter distilled model, service providers can rapidly build toolchains for software, cybersecurity, office productivity, and other verticals.

The MiMo-V2.6 series' low price and high performance present platforms with both a developer-traffic opportunity and a compliance challenge.

1. Platform integration: The OpenCode developer platform has launched the series, with the Flash version free for one week to attract trials. Platforms can follow suit by launching large models to expand their developer ecosystems.

2. Operations and cost: Flash's API pricing is among the world's two lowest for frontier models, and its cached-input rate is ultra-low, making it a strong candidate as a platform default model or for long-context tasks. Multimodal and long-context support also enable complex platform applications.

3. Risk mitigation: An Anthropic report alleges that Xiaomi and other Chinese AI labs engaged in "illegal distillation" of Claude's capabilities. Platforms should establish compliance review and risk-control mechanisms before integrating Xiaomi models, and track follow-up responses and regulatory developments.

4. Recruitment and management reference: Xiaomi promotes its models through open weights, free trials, and a community platform. Platforms can borrow this playbook to run developer activities in partnership with model vendors.

MiMo-V2.6 offers rare technical detail on large-scale reinforcement learning training, along with open resources, and it reignites debates over model copyright and compliance.

1. Industry development: For the first time, an open-source model has topped an aggregate intelligence index and matched leading closed-source models, indicating that post-training reinforcement learning is now the primary driver of capability gains. The fact that low-cost, large-scale training was achieved entirely on domestic chips is also noteworthy.

2. Technical methods: Training uses the You Only RL Once strategy, integrating multi-domain tasks and lightweight agent scaffolding into a single training run, combined with asynchronous group-relative policy optimization to improve GPU utilization. A dedicated "cheating detection agent" and a frozen routing mechanism ensure reward reliability, with cheating trajectories accounting for less than 2% of samples.

3. Open resources: A paper-level technical report, an end-to-end training framework, more than 7,000 auto-graded tasks, and a 9-billion-parameter distilled model have all been released, supporting reproduction and further research.

4. Controversy and legal issues: Anthropic alleges that Xiaomi routed user conversations into Claude via the OpenClaw and OpenCode tools, generating around 400,000 interaction rounds and amounting to "illegal distillation." The episode raises questions about training-data provenance, terms-of-service compliance, and cross-border AI governance that warrant policy attention.

Disclaimer: The "Quick Summary" content is entirely generated by AI. Please exercise discretion when interpreting the information. For issues or corrections, please email run@ebrun.com .

I am a Brand Seller Factory Service Provider Marketplace Seller Researcher Read it again.

2026年9月22日,小米正式发布MiMo-V2.6系列大模型。第三方分析机构Artificial Analysis发布的最新智能指数测评显示,该系列旗舰款MiMo-V2.6-Pro得分达46分,位列当前公开可用模型首位,追平同日发布的xAI Grok 4.7,超过Kimi K3、Qwen、DeepSeek V4.1系列等开源竞品,也高于Grok 4.6、谷歌Gemini 3.8 Flash等闭源模型。目前Claude Opus 5、OpenAI GPT-5.6 Sol等闭源模型仍在部分细分测评中保持领先,MiMo-V2.6-Pro并未实现全基准覆盖。

MiMo-V2.6-Pro为混合专家架构模型,总参数1.02万亿,单请求激活参数420亿,支持100万token上下文窗口,可接收文本、图像、音频、视频多模态输入,常规输出速度约134token每秒。其API调用定价为每百万无缓存输入token0.435美元、每百万输出token0.87美元,折算单智能指数测试任务成本约0.13美元,处于智能水平与成本的帕累托最优区间。用户可免费从Hugging Face平台下载模型权重,基于MIT许可自行定制、微调,部署至自有或租赁硬件用于生产流程。同性能级闭源模型API定价普遍高出数倍至数十倍,Claude Opus 5每百万输入输出总价达30美元,GPT-5.6 Sol标准模式总价达35美元。

同系列的MiMo-V2.6-Flash定位高并发生产场景,总参数3100亿,单请求激活参数150亿,同样支持百万token上下文窗口与多模态输入能力,多项长流程智能体基准测试得分与Pro版本差距在5分以内,在CyberGym网络安全测评中得分甚至超过Pro。其API定价为每百万无缓存输入token0.14美元、每百万输出token0.28美元,缓存命中输入价格低至每百万token0.0028美元,是当前全球定价第二低的主流前沿大模型API。同步推出的MiMo-V2.6-Pro-UltraSpeed版本,输出速度最高可达普通Pro版的20倍。

小米2025年开始公开扩充MiMo大模型产品线,2026年4月推出的V2.5系列已确立稀疏混合专家架构、百万token上下文、宽松许可、低定价等核心特性,6月上线开源终端编码智能体MiMo Code与自适应智能体框架HarnessX,本次V2.6系列将此前积累的脚手架、智能体能力整合进了模型训练流程。

小米披露该版本性能跃升主要来自大规模强化学习训练。整个训练流程耗时不到6天,Pro版本训练总成本约262万美元,Flash版本约85万美元,其中超五成成本用于生成训练轨迹与结果评分。训练采用“You Only RL Once”策略,将编码、专业办公、视觉任务、网络安全等多领域任务与多种轻量智能体脚手架整合进单次训练流程,搭配异步组相对策略优化算法提升GPU利用率,避免长任务拖慢整体训练节奏。为防范模型通过漏洞刷取奖励而非真正完成任务,小米搭建了两套群组评分机制,训练前专门部署“作弊检测智能体”排查环境漏洞,冻结混合专家路由机制减少训练漂移,最终确认的奖励作弊轨迹占比低于2%。训练完成后,Pro版本在DeepSWE编码测试中的得分从V2.5版本的58.4升至72.6,Flash版本从48.8升至65.7。

除模型权重外,小米同步开放了完整强化学习技术报告、端到端训练框架、7000余个带自动评分器的现成训练任务,覆盖软件开发、网络安全、办公、网页设计、音乐创作等领域,同时放出基于MiMo强化学习轨迹蒸馏的90亿参数小模型,供开发者复现、扩展相关工作。官方演示内容显示,该系列模型可协调多个智能体完成可玩3D世界生成、Blender场景搭建、桌面软件操作、机械臂视觉控制、前端设计、音视频制作等长流程任务,内部测试中已支撑科研人员完成PFAS吸附材料筛选、数学定理形式化验证等专业工作。

小米MiMo团队负责人、前DeepSeek研究员罗福利在社交平台提及,V2.6是开源模型团队迄今规模最大的单次强化学习训练之一,团队投入数十人,相关研发与工程挑战超过其参与DeepSeek R1研发时的水平。开源AI知识工作初创公司Paper Instruments首席执行官Daanish Khazi在社交平台发文,称小米完全基于国产芯片完成本次训练,以相对较低的算力成本实现了大幅能力提升,相关成果将推动研究者更新后训练缩放定律的认知。卡内基梅隆大学计算机科学教授、大模型量化工具bitsandbytes创作者Tim Dettmers在社交平台发文,称Flash版本是3000亿至5500亿参数级中的最优模型,表现优于DeepSeek V4.1、GLM 5.3 Flash。开发者平台OpenCode已上线该系列模型,Flash版本对用户免费开放一周。

本次发布两周前,Anthropic发布威胁情报报告,将小米列入2025年12月至2026年8月期间滥用Claude模型的七家中国AI实验室名单,其余六家包括阿里巴巴、月之暗面、DeepSeek、智谱、MiniMax、商汤。报告称这些机构累计生成约1.9亿轮对话,通过“非法蒸馏”手段抽取Claude能力训练自有模型。其中涉及小米的案例显示,2026年3至4月的20天里,小米通过OpenClaw、OpenCode工具将自有MiMo模型的用户对话、编码会话内容传入Claude,累计产生超40万轮交互,用于丰富后续模型的训练数据。报告将该行为与本次MiMo-V2.6-Pro登顶开源榜单相关联,截至发稿小米尚未就该指控作出公开回应。

本文首发于 亿邦动力 官方网站

文章来源:亿邦动力

广告
微信
朋友圈

FAQ回顾

小米MiMo-V2.6是什么?

小米于2026年9月22日发布MiMo-V2.6系列大模型,包含旗舰款MiMo-V2.6-Pro、面向高并发生产场景的MiMo-V2.6-Flash及高速版UltraSpeed。Pro版总参数1.02万亿,单请求激活420亿,支持100万token上下文和多模态输入;Flash版总参数3100亿,单请求激活150亿,同样具备百万token上下文和多模态能力。

MiMo-V2.6-Pro在性能榜单上表现如何?

在Artificial Analysis智能指数测评中,MiMo-V2.6-Pro得分46分,位列公开可用模型首位,追平xAI Grok 4.7,超越DeepSeek V4.1、Qwen、Kimi K3等开源模型,也高于Grok 4.6、Gemini 3.8 Flash等闭源模型;不过Claude Opus 5、GPT-5.6 Sol等部分细分测评仍领先。

MiMo-V2.6的API定价是多少?

MiMo-V2.6-Pro每百万无缓存输入token 0.435美元、输出token 0.87美元;Flash版分别为0.14美元和0.28美元,缓存命中输入低至0.0028美元,为全球第二低的主流前沿大模型API。相比之下Claude Opus 5总价30美元,GPT-5.6 Sol总价35美元,MiMo价格优势显著。

小米MiMo-V2.6为什么训练成本低?

Pro版训练总成本约262万美元,Flash约85万美元,耗时不到6天。这得益于“You Only RL Once”策略,把编码、办公、视觉、网络安全等任务与轻量智能体脚手架整合进单次强化学习流程,并采用异步组相对策略优化提升GPU利用率,还通过作弊检测智能体机制确保训练质量。

MiMo-V2.6支持哪些应用场景?

MiMo-V2.6支持文本、图像、音频、视频多模态输入,可协调多个智能体完成3D世界生成、Blender场景搭建、桌面软件操作、机械臂视觉控制、前端设计、音视频制作等长流程任务。内部测试中已支撑科研人员完成PFAS吸附材料筛选、数学定理形式化验证等专业工作。

这么好看,分享一下?

朋友圈 分享

APP内打开

+1
+1
微信好友 朋友圈 新浪微博 QQ空间
关闭
收藏成功
发送
/140 0