小米深夜上线三大自研MiMo-V2系列模型:旗舰Pro、全模态基座Omni与TTSXiaomi Launches Three Self-Developed MiMo-V2 Series Models Late at Night: Flagship Pro, Full-Modal Base Omni, and TTS


AI free token/AI免费积分
header
关注

转自,转载IT之家

3 月 19 日消息,今天凌晨,小米宣布推出三款大模型 —— MiMo-V2-Pro & Omni & TTS。当下,这些模型已经登陆 Xiaomi miclaw、MiMo Studio、金山办公、小米浏览器,借助 OpenClaw、OpenCode、KiloCode、Blackbox、Cline 接入,可限时免费感受一周。

IT之家汇总官方对各模型的介绍如下:

小米面向 Agent 时代的旗舰基座模型:Xiaomi MiMo-V2-Pro

Xiaomi MiMo-V2-Pro 专为现实世界中高强度的 Agent 工作情形而构建。其具备超过 1T 的总参数量(42B 激活参数),使用创新的混合注意力架构,并兼容 1M 超长上下文长度。小米还在更为广泛的 Agent 情形中持续 Scaling 算力,进一步拓展了智能的动作空间,达成了从 Coding 到 Claw 的关键泛化。在 Artificial Analysis 排行榜上,MiMo-V2-Pro 位列全球第八,国内第二。

在 OpenClaw、Claude Code 等智能体框架中,MiMo-V2-Pro 可以在无人工干预的条件下完成复杂工作流编排、长程规划与精准工具调用,并持续可靠地交付最终结果。总体使用体感已超越 Claude Sonnet 4.6,逼近 Opus 4.6,但模型 API 定价仅为其 1/5,降低了前沿智能的使用门槛。

MiMo-V2-Pro 专为 Agent 情形深度优化。其针对复杂多样的 Agent Scaffold 开展 SFT & RL,具备更强的工具调用与多步推理能力。在 OpenClaw 标准实测榜单 PinchBench、ClawEval 上,MiMo-V2-Pro 成效处于全球顶尖。同时,凭借 1M 上下文窗口,MiMo-V2-Pro 可以从容支撑高强度的真实 Claw 复杂应用流。

MiMo-V2-Pro 可以参与更严肃的代码工程构建。内部工程师实测表明,其体感已接近 Claude Opus 4.6,并展现出高阶的代码智能:具备更出色的平台外观与任务规划能力、更优雅的代码风格,以及更高质直接的难题应对路径。

MiMo-V2-Pro 模型现已正式开放 API 服务,兼容 1M 上下文长度,并依据使用量分段计价:

  • 256K 上下文以内:输入 $1 / 百万 tokens,输出 $3 / 百万 tokens

  • 1M 上下文以内:输入 $2 / 百万 tokens,输出 $6 / 百万 tokens

访问 https://platform.xiaomimimo.com,即刻接入 API。

MiMo Claw 模块现已全面打通金山 WebOffice 生态。原生兼容 Word、Excel、PPT、PDF 四大主流格式,无缝覆盖超 95% 的日常文档类型。

Xiaomi MiMo 底层推理引擎与金山办公生态达成框架级集成。WPS 灵犀现已接入 MiMo-V2-Pro 模型,向灵犀 Claw 提问或布置任务,办公更高质。

小米面向 Agent 时代的全模态基座模型:Xiaomi MiMo-V2-Omni

其专为现实世界中复杂的多模态交互与执行情形而生,可无缝接入各种 Agent 框架,达成了从理解到操控的跨越,大幅降低了全模态 Agent 的落地门槛。

音频理解维度,兼容从环境声分类、多说话人分离、音频-视觉联合推理,到超过 10 小时连续长音频的深度理解。综合表现超越 Gemini 3 Pro,是当前最强的音频理解基座模型之一。

图像理解维度,MiMo-V2-Omni 展现出强大的多学科视觉推理与复杂图表分析能力,超越 Claude Opus 4.6,逼近 Gemini 3 Pro 等顶尖闭源模型水平。

视频理解维度,兼容原生音视频联合输入,达成真正的多模态视频理解。借助创新的视频预训练,模型具备强大的情境感知与将来推理能力。

MiMo-V2-Omni 可以跨模态理解复杂环境、自主制定并执行计划、在遇到异常时实时修正策略,最终端到端地交付完整结果。

MiMo-V2-Omni 模型现已正式开放 API 服务,兼容 256K 上下文长度,输入 $0.4 / 百万 tokens,输出 $2 / 百万 tokens。

访问 https://platform.xiaomimimo.com,即刻接入 API。

另外,MiMo-V2-Omni 联合 OpenClaw、OpenCode、KiloCode、Blackbox 及 Cline 等五大 Agent 开发框架团队,为全球开发者给出为期一周的限时免费接口兼容。

专为 Agent 时代的全模态交互而生:Xiaomi MiMo-V2-TTS 语音合成大模型

Xiaomi MiMo-V2-TTS 是小米自主研发的语音合成大模型。它基于自研 Audio Tokenizer 和多码本语音-文本联合建模架构,经过上亿小时语音数据的大规模预训练与多维度强化学习,达成了高度可控的多粒度语音风格控制。MiMo-V2-TTS 兼容从总体风格定调到局部情绪表达的精准调节,能在同一句话内完成语气转折和情感递变;真实还原人类说话的自然韵律;在唱歌时,也能准确表达音高和节奏,自然且富有表现力。

在训练过程中,MiMo-V2-TTS 首先借助超大规模语音-文本混合预训练,在海量数据中习得了强大的跨模态对齐与理解生成的统一能力;在此基础上,借助少量高质量监督数据的微调,模型取得了可泛化的多粒度与多风格指令控制能力。

为进一步激发模型在大规模预训练中积累的高表现力语音生成潜力,小米引入了多维度强化学习,兼顾了稳定性与表现力。具体而言,MiMo-V2-TTS 在强化学习阶段,围绕更自然的韵律、更稳定的音质、更准确的字词表达、更高质量的音色克隆以及不同情形下恰当的语气和表达方式等多个维度持续优化。得益于多层码本建模架构,模型在高保真的离散 token 空间中对语音开展建模,充分保留了原始语音中的多元信息,使强化学习阶段可以直接利用语音相关奖励信号对模型开展优化,从而让多维奖励信号更有效地作用于生成过程。

MiMo-V2-TTS 兼容从总体到局部的多层次语音风格控制。使用者可借助自然语言指令设定总体语音基调,同时对句内局部片段开展细粒度的情绪调节,达成同一语句中语气转折与情感递变的自然过渡。

该模型还具备多元的多元表达能力:兼容多种方言的自然发音,可开展角色扮演式的风格化演绎,更能达成高质量的歌声合成 —— 让同一个模型既能说、能演、也能唱,兼容东北话、四川话、河南话、粤语、台湾腔等多种方言。

广告声明:文内含有的对外跳转链接(涵盖不限于超链接、二维码、口令等形式),用于传递更多信息,节省甄选时间,结果仅供参考,IT之家所有文章均包含本声明。

这些回复亮了

discusser-avatar
天下为公1233:华为没有发布过声明,说小米抄袭的;同样,阿里也没有发布过声明,说华为抄袭自己。所谓的华为盘古抄袭,是水军捏造的。从来没见过有正规的官方媒体报道,也没有国家机关处理过,连阿里自己都没说过华为抄袭自己的。这些所谓的消息来源,都是某某论坛某某帖子,谁都可以发布。不信?现在就可以跟我对线,是不是某某论坛某某帖子来的?那么什么事正规官方媒体报道呢?我举个例子----新华社:小米销售方式没有底线收起

盘古又不是抄袭又那么牛逼,为什么团队解散了? 但凡有点思维逻辑都明白,事实就是抄袭qianwen模型,行内人都锤爆了

discusser-avatar
天下为公1233:你所谓的团队解散的官方报道在哪里?XX论坛XX帖子是吧?https://www.huaweicloud.com/product/huaweicloudstack/hclargeaimodels.html这是华为官网的机型介绍页面,团队解散了,这个大模型的机型是谁在研发销售维护?哦,人工智能,大模型自己就可以维护自己,那华为的大模型的确厉害,都可以自己维护自己,不需要人工团队了,毕竟“人工团队”在黑子口中已经解散了嘛!都解散了,可不是大模型自己维护自己?收起

肯定是加拿大人在维护啊,毕竟中专毕业,肯定最懂大模型了,狠狠研发加拿大大house和脚铐

GIF
discusser-avatar
天下为公1233:华为没有发布过声明,说小米抄袭的;同样,阿里也没有发布过声明,说华为抄袭自己。所谓的华为盘古抄袭,是水军捏造的。从来没见过有正规的官方媒体报道,也没有国家机关处理过,连阿里自己都没说过华为抄袭自己的。这些所谓的消息来源,都是某某论坛某某帖子,谁都可以发布。不信?现在就可以跟我对线,是不是某某论坛某某帖子来的?那么什么事正规官方媒体报道呢?我举个例子----新华社:小米销售方式没有底线收起

呵呵 华为一点难题都不能说么

本文来源:https://voice.hupu.com/bbs/637975395
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

header

Wembanyama DPOYbadge

03-19 08:14

Follow

Reposted from IT Home

March 19 news, early this morning, Xiaomi announced the launch of three large models - MiMo-V2-Pro & Omni & TTS. These models have now landed on Xiaomi miclaw, MiMo Studio, Kingsoft Office, and Xiaomi Browser, and can be accessed via OpenClaw, OpenCode, KiloCode, Blackbox, and Cline, with a limited-time free trial for one week.

IT Home summarizes the official introduction of each model as follows:

Xiaomi's flagship foundation model for the Agent era: Xiaomi MiMo-V2-Pro

Xiaomi MiMo-V2-Pro is built for high-intensity Agent work scenarios in the real world. It has over 1T total parameters (42B active parameters), uses an innovative hybrid attention architecture, and supports an ultra-long 1M context length. Xiaomi is also continuously scaling compute for broader Agent scenarios, further expanding the intelligent action space and achieving key generalization from Coding to Claw. On the Artificial Analysis leaderboard, MiMo-V2-Pro ranks 8th globally and 2nd domestically.

In agent frameworks such as OpenClaw and Claude Code, MiMo-V2-Pro can complete complex workflow orchestration, long-horizon planning, and precise tool calling without human intervention, while reliably delivering final results. The overall user experience already surpasses Claude Sonnet 4.6 and approaches Opus 4.6, but the model's API pricing is only 1/5 of it, lowering the barrier to frontier intelligence.

MiMo-V2-Pro is deeply optimized specifically for Agent scenarios. It undergoes SFT & RL across diverse and complex Agent scaffolds, with stronger tool-calling and multi-step reasoning capabilities. On OpenClaw's standard benchmark leaderboards PinchBench and ClawEval, MiMo-V2-Pro achieves world-leading results. Also, with its 1M context window, MiMo-V2-Pro can easily support high-intensity real-world Claw complex application flows.

MiMo-V2-Pro can participate in more serious software engineering work. Internal engineer testing shows that its feel is close to Claude Opus 4.6, and it demonstrates advanced code intelligence: better platform appearance and task planning capabilities, more elegant code style, and higher-quality, more direct approaches to difficult problems.

The MiMo-V2-Pro model now officially offers API services, supports 1M context length, and is priced in tiers based on usage:

Within 256K context: input $1 / million tokens, output $3 / million tokens

Within 1M context: input $2 / million tokens, output $6 / million tokens

Visit https://platform.xiaomimimo.com to connect to the API now.

The MiMo Claw module is now fully integrated with Kingsoft WebOffice's ecosystem. It natively supports the four major formats of Word, Excel, PPT, and PDF, seamlessly covering over 95% of everyday document types.

Xiaomi MiMo's underlying inference engine has achieved framework-level integration with the Kingsoft Office ecosystem. WPS Lingxi now supports the MiMo-V2-Pro model. Ask Lingxi Claw questions or assign tasks, and office work becomes more efficient.

Xiaomi's all-modal foundation model for the Agent era: Xiaomi MiMo-V2-Omni

It is built for complex multimodal interaction and execution scenarios in the real world, can seamlessly connect to various Agent frameworks, and achieves the leap from understanding to manipulation, greatly lowering the barrier to deployment for all-modal Agents.

In audio understanding, it supports everything from environmental sound classification, multi-speaker separation, and audio-visual joint reasoning, to deep understanding of continuous long audio exceeding 10 hours. Its overall performance surpasses Gemini 3 Pro, making it one of the strongest audio understanding foundation models today.

In image understanding, MiMo-V2-Omni demonstrates strong multidisciplinary visual reasoning and complex chart analysis capabilities, surpassing Claude Opus 4.6 and approaching top closed-source models such as Gemini 3 Pro.

In video understanding, it supports native joint audio-video input, achieving true multimodal video understanding. Through innovative video pretraining, the model has strong situational awareness and future reasoning capabilities.

MiMo-V2-Omni can understand complex environments across modalities, autonomously formulate and execute plans, adjust strategies in real time when anomalies occur, and ultimately deliver complete results end-to-end.

The MiMo-V2-Omni model now officially offers API services, supports 256K context length, with input at $0.4 / million tokens and output at $2 / million tokens.

Visit https://platform.xiaomimimo.com to connect to the API now.

In addition, MiMo-V2-Omni, together with five major Agent development framework teams including OpenClaw, OpenCode, KiloCode, Blackbox, and Cline, offers global developers a limited-time one-week free interface compatibility.

Built for all-modal interaction in the Agent era: Xiaomi MiMo-V2-TTS speech synthesis large model

Xiaomi MiMo-V2-TTS is a speech synthesis large model independently developed by Xiaomi. It is based on a self-developed Audio Tokenizer and a multi-codebook speech-text joint modeling architecture. After large-scale pretraining on hundreds of millions of hours of speech data and multi-dimensional reinforcement learning, it achieves highly controllable multi-granularity speech style control. MiMo-V2-TTS supports precise adjustment from overall style setting to local emotional expression, and can complete tone shifts and emotional transitions within the same sentence; it faithfully reproduces the natural prosody of human speech; and when singing, it can accurately express pitch and rhythm, naturally and expressively.

During training, MiMo-V2-TTS first leverages ultra-large-scale speech-text mixed pretraining to learn strong cross-modal alignment and unified understanding-generation capabilities from massive data. On this basis, with fine-tuning on a small amount of high-quality supervised data, the model acquires generalizable multi-granularity and multi-style instruction control capabilities.

To further unlock the highly expressive speech generation potential accumulated during large-scale pretraining, Xiaomi introduced multi-dimensional reinforcement learning, balancing stability and expressiveness. Specifically, during the reinforcement learning stage, MiMo-V2-TTS continuously optimizes across multiple dimensions, including more natural prosody, more stable sound quality, more accurate word and character expression, higher-quality voice cloning, and appropriate tone and expression in different scenarios. Thanks to the multi-layer codebook modeling architecture, the model models speech in a high-fidelity discrete token space, fully preserving the diverse information in the original speech, allowing the reinforcement learning stage to directly use speech-related reward signals to optimize the model, thereby making multi-dimensional reward signals act more effectively on the generation process.

MiMo-V2-TTS supports multi-level speech style control from overall to local levels. Users can set the overall speech tone through natural language instructions while performing fine-grained emotional adjustment on local segments within a sentence, achieving natural transitions of tone and emotional shifts within the same utterance.

The model also has diverse expressive capabilities: it supports natural pronunciation of multiple dialects, can perform role-play style stylized performances, and can even achieve high-quality singing synthesis - letting the same model speak, perform, and sing, with support for multiple dialects including Northeastern Mandarin, Sichuanese, Henan dialect, Cantonese, and Taiwanese accent.

Advertisement statement: The external jump links contained in this article (including but not limited to hyperlinks, QR codes, command phrases, etc.) are used to convey more information and save screening time. The results are for reference only. All IT Home articles include this statement.

These replies are popular

discusser-avatar

Tianxiaweigong1233: Huawei has never issued a statement saying Xiaomi plagiarized; likewise, Alibaba has never issued a statement saying Huawei plagiarized it. The so-called Huawei Pangu plagiarism was fabricated by water armies. No legitimate official media has ever reported it, and no state agency has ever handled it. Even Alibaba itself has never said Huawei plagiarized it. These so-called sources all come from some forum or some post, which anyone can publish. Don't believe it? You can come and confront me right now - isn't it from some forum or some post? Then what counts as legitimate official media reporting? Let me give an example - Xinhua News Agency: Xiaomi's sales methods have no bottom line Collapse

If Pangu wasn't plagiarized and was so awesome, why was the team disbanded? Anyone with basic logic can understand that the truth is it plagiarized the Qianwen model. People in the industry have thoroughly exposed it.

discusser-avatar

Tianxiaweigong1233: Where is the official report about the team being disbanded that you mention? Some forum and some post, right? https://www.huaweicloud.com/product/huaweicloudstack/hclargeaimodels.html This is the model introduction page on Huawei's official website. If the team was disbanded, who is researching, selling, and maintaining this large model's product line? Oh, artificial intelligence - the large model can maintain itself. Then Huawei's large model really is impressive. It can maintain itself without needing a human team. After all, in the mouths of detractors, the "human team" has already been disbanded! If they're all disbanded, doesn't that mean the large model maintains itself? Collapse

It must be the Canadians maintaining it. After all, with only a vocational school diploma, they surely understand large models best - vigorously developing Canadian mansions and ankle shackles.

GIF

discusser-avatar

Tianxiaweigong1233: Huawei has never issued a statement saying Xiaomi plagiarized; likewise, Alibaba has never issued a statement saying Huawei plagiarized it. The so-called Huawei Pangu plagiarism was fabricated by water armies. No legitimate official media has ever reported it, and no state agency has ever handled it. Even Alibaba itself has never said Huawei plagiarized it. These so-called sources all come from some forum or some post, which anyone can publish. Don't believe it? You can come and confront me right now - isn't it from some forum or some post? Then what counts as legitimate official media reporting? Let me give an example - Xinhua News Agency: Xiaomi's sales methods have no bottom line Collapse

Haha, so Huawei can't take any criticism at all?

Download the Hupu app to see more popular replies

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry