1T参数、订阅14.9元起,小米MiMo V2.5正面硬刚旗舰1T parameters, subscription from 14.9 yuan, Xiaomi MiMo V2.5 takes on flagships head-on


AI free token/AI免费积分

本文来源:https://mp.weixin.qq.com/s/inERC-QaBpzApf-0AtPEdw
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

1T参数、订阅14.9元起,小米MiMo V2.5正面硬刚旗舰

一直在大模型竞争里"慢半拍"的小米,这次把旗舰参数直接拉满。MiMo-V2.5系列全新发布,旗舰款MiMo-V2.5-Pro做到1T总参数、42B激活、1M超长上下文,官方直言在高强度智能体场景下,表现媲美Claude Opus 4.6。

旗舰款1T参数,UltraSpeed跑出1000 tokens/s

这代MiMo的产品矩阵,照着"既要大又要快"的思路铺开。

旗舰MiMo-V2.5-Pro主打极致Agent性能,1T总参数、42B激活、1M上下文,对标的是国内外顶配旗舰。更猛的是MiMo-V2.5-Pro-UltraSpeed,靠算法和系统两层优化——FP4混合量化无损压缩体积,配合DFlash并行投机解码,TileRT系统把计算与数据搬运极致重叠,峰值速度冲到1000 tokens每秒,目前开放内测。

还有原生全模态的MiMo-V2.5,支持图像、视频、音频、文本的跨模态理解;以及精品音色的TTS系列(限时免费)和能识别中英双语加方言的ASR系列。

对普通用户来说,这意味着MiMo不只回答文字,还能听、能说。TTS能精细控制语速、情绪和语气,ASR按0.5元每小时的音频计费。从"看得到的AI"到"听得见、会开口的AI",小米补的是全模态这块短板。

需要注意一个时间节点:MiMo-V2系列已于2026年6月30日下线,原模型名称失效,需要切到V2.5系列。小米这一刀,是把老版本连根换掉,逼着开发者上新车。

定价走订阅+按量,不是"买断"那一套

很多人关心这代怎么收费。需要说清楚的是,MiMo走的是订阅和按量计费两条线,不是坊间传的"一次性买断"。

模型API按token计价:MiMo-V2.5-Pro输入缓存命中0.025元每百万tokens、未命中3元,输出6元;轻量的MiMo-V2.5输入未命中1元、输出2元。

面向C端,MiMo Claw正式版限时特惠14.9元每月,把旗舰mimo-v2.5-pro、OpenClaw原生适配和金山办公生态打包,TokenPlan还能包月或包年订阅,免费体验时长也从每天1小时升级到4小时。

把旗舰模型和金山办公生态绑在一起,瞄准的是文档、表格、PPT这些每天都要用的办公场景——这才是订阅用户愿意持续付费的理由,而不是模型参数本身。

ASR按音频时长计费0.5元每小时。整套定价的逻辑是"用多少花多少",加上订阅封顶,照顾开发者和重度用户两头。

小米补上大模型这块,硬件才算有灵魂

从2017年起,小米在AI上没断过投入——小爱同学的语音、手机端AI影像、汽车智能驾驶。但自研大模型这条线,直到MiMo才算真正补齐。

手机、汽车、眼镜跑得再齐,缺一个自研大模型当大脑,整套生态就是硬件多、灵魂少。

MiMo补的就是这个灵魂。你觉得,小米这套"1T参数+14.9元订阅"的组合拳,能在拥挤的大模型市场里挤进第一梯队吗?评论区聊聊👇

关注我,下一篇继续聊国产大模型的新进展;觉得有用,点个「在看」👇

 

Source: https://mp.weixin.qq.com/s/inERC-QaBpzApf-0AtPEdw If there are any copyright issues, please email: 13951431913#139.com or submit a comment on the right.

1T parameters, subscription starting from 14.9 yuan, Xiaomi MiMo V2.5 goes head-to-head with flagships

Xiaomi, which has always been "half a step behind" in the large model competition, this time maxes out the flagship parameters directly. The MiMo-V2.5 series is newly launched, with the flagship MiMo-V2.5-Pro featuring 1T total parameters, 42B activated, and 1M ultra-long context. The official statement claims that in high-intensity agent scenarios, its performance rivals Claude Opus 4.6.

Flagship 1T parameters, UltraSpeed reaches 1000 tokens/s

This generation of MiMo product matrix is laid out following the logic of "both big and fast".

The flagship MiMo-V2.5-Pro focuses on ultimate Agent performance, with 1T total parameters, 42B activated, and 1M context, benchmarking against top-tier domestic and international flagships. Even more impressive is the MiMo-V2.5-Pro-UltraSpeed, which relies on two-layer optimization of algorithms and systems—FP4 hybrid quantization losslessly compresses volume, coupled with DFlash parallel speculative decoding, and the TileRT system overlaps computation and data movement to the extreme, pushing peak speed to 1000 tokens per second, currently open for internal testing.

There is also the natively full-modal MiMo-V2.5, supporting cross-modal understanding of images, videos, audio, and text; as well as the TTS series with premium tones (free for a limited time) and the ASR series capable of recognizing Chinese-English bilingual plus dialects.

For ordinary users, this means MiMo can not only answer text but also listen and speak. TTS can finely control speech rate, emotion, and tone, while ASR charges based on 0.5 yuan per hour of audio. From "visible AI" to "audible and speaking AI", Xiaomi is filling the gap in full-modal capabilities.

Note a critical time node: The MiMo-V2 series was taken offline on June 30, 2026, and the original model names are invalid, requiring a switch to the V2.5 series. Xiaomi's move is to replace the old version completely, forcing developers to upgrade.

Pricing follows subscription + pay-as-you-go, not the "buyout" model

Many are concerned about how this generation is charged. It needs to be clarified that MiMo follows two lines of subscription and pay-as-you-go billing, not the "one-time buyout" rumored in the industry.

Model API is priced by token: MiMo-V2.5-Pro input cache hit is 0.025 yuan per million tokens, miss is 3 yuan, output is 6 yuan; the lightweight MiMo-V2.5 input miss is 1 yuan, output is 2 yuan.

For the C-end, the MiMo Claw formal version has a limited-time special offer of 14.9 yuan per month, bundling the flagship mimo-v2.5-pro, OpenClaw native adaptation, and the Kingsoft Office ecosystem. TokenPlan can also be subscribed monthly or yearly, and the free trial duration has been upgraded from 1 hour per day to 4 hours.

Bundling the flagship model with the Kingsoft Office ecosystem targets daily office scenarios such as documents, spreadsheets, and PPTs—this is the reason subscription users are willing to pay continuously, not the model parameters themselves.

ASR charges by audio duration at 0.5 yuan per hour. The logic of the entire pricing scheme is "pay as you use", plus a subscription cap, catering to both developers and heavy users.

Xiaomi fills the large model gap, giving hardware its soul

Since 2017, Xiaomi has not stopped investing in AI—Xiao Ai's voice, mobile AI imaging, and automotive intelligent driving. However, the self-developed large model line was only truly completed with MiMo.

Whether it's phones, cars, or glasses running in sync, without a self-developed large model as the brain, the entire ecosystem has more hardware and less soul.

MiMo fills this soul. Do you think Xiaomi's combination of "1T parameters + 14.9 yuan subscription" can squeeze into the first tier in the crowded large model market? Discuss in the comments 👇

Follow me, next article will continue discussing new progress in domestic large models; if useful, click "Like" 👇

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