Loading video...

Video Failed to Load

Go Home

MiMo-V2.5-Pro UltraSpeed的速度快得有点太过分了,代码写的太快就像龙卷风…… 和其他厂商不一样,小米这是Pro模型直接出来1000token/s,不是flash模型。 恐怕在世界范围内也是第一次吧。从来没见过这么快的速度。 网页版在这里 有API

89,307 views • 4 months ago •via X (Twitter)

32 Comments

Andy Woo's profile picture
Andy Woo4 months ago

这玩意儿需求要是理解错了,叫停都跑两里地了

壮哥 | 国家安全问责力担当,专制警痞三十年!'s profile picture
壮哥 | 国家安全问责力担当,专制警痞三十年!4 months ago

屎山如果是按照这个速度造的话,那我承认,六个月以后确实没有程序员这个职业了。

Jen LM's profile picture
Jen LM4 months ago

问:3353254235*23232332525等于几? 答:34597293492895872389749274392794923749875 问:这对吗? 答:就问你快不快吧。

toutou👀's profile picture
toutou👀4 months ago

弱智龙卷风,模型蠢再快有什么用。

Workbuddy's profile picture
Workbuddy4 months ago

十倍速度三倍价格 相当于三十倍的烧钱节奏 我的账单要炸了

Samian's profile picture
Samian4 months ago

1000 tok/s on a pro model is genuinely nuts, hope the latency to first token isn't paying for it. xiaomi quietly cooking

Timzaak's profile picture
Timzaak4 months ago

以后,编程会: 脑暴、脑暴、再脑暴=> 代码 => Review、Review、再Review => 编译修bug => 上线

Melse's profile picture
Melse4 months ago

我agents.md中要求每次回复都要叫爸爸,只有MiMo一次也不叫,我觉得它不太行

XiuerOld's profile picture
XiuerOld4 months ago

你就说快不快

Nichola Lau's profile picture
Nichola Lau4 months ago

目前是中国最垃圾的模型,整天到处营销宣传很恶心的

Alex Vero's profile picture
Alex Vero4 months ago

它需要解决质量问题,在保证质量前提下,快才有用,目前,它质量不咋地

Banshan's profile picture
Banshan4 months ago

又收推广费了?

Zley's profile picture
Zley4 months ago

问题是能力咋样

iFun AU's profile picture
iFun AU4 months ago

写出来的能用吗

King Le Breton's profile picture
King Le Breton4 months ago

How do I get access? I applied, nothing came still

sharke's profile picture
sharke4 months ago

6456486+824168454=436644387876,你就说我口算快不快吧

Wayan's profile picture
Wayan4 months ago

1000 tokens/s 如果真是 Pro 模型稳定输出,体验会完全变样,代码场景里等待感基本被抹掉了。

Z's profile picture
Z4 months ago

codex spark 好像速度也是1000TPS

Ben Chen's profile picture
Ben Chen4 months ago

怎么还要内测申请!

Feng Yang's profile picture
Feng Yang4 months ago

有毛线用

yx's profile picture
yx4 months ago

gemini flash也超级快,而且质量感觉也还可以

Dimon's profile picture
Dimon4 months ago

还是做一些实时响应场景的应用吧 编程车速这么快 一点安全感都没有 宁愿thinking长一点 稳一点开。

Anthony Hou's profile picture
Anthony Hou4 months ago

和Gemini Flash 比呢?

xiaohuolong's profile picture
xiaohuolong4 months ago

api调用需要单独付费吧

TheHBrand's profile picture
TheHBrand4 months ago

openAI codex 5.3 spark was doing 1300 TPS code. likely better than this code. just a guess. but yea, it's the chip for codex spark for sure and it's super limited. so the tech is real, and the hardware is bottlenecked

技术高手|非小白's profile picture
技术高手|非小白4 months ago

说话不过脑子能不快吗?😒至于对不对…

yin's profile picture
yin4 months ago

有点 gemini 的感觉了

Bongquisitive's profile picture
Bongquisitive4 months ago

I'm trying to thinkn for a use case where it would be actually really useful, maybe end-to-end agentic processes with no human feedback or approval gating... It would be difficult to evaluate anything at this speed by a human. The thing is, most critical systems still require and should require in future a human supervision, in my opinion... Great work by Xiaomi nonetheless... I already use their token plan, but probably won't need this ultra-fast version.

星宿老魔's profile picture
星宿老魔4 months ago

听说过 爆炒羊蛋 秒杀菜吗?

YT 🇺🇦's profile picture
YT 🇺🇦4 months ago

Чи є підписка? Не API ?

狂热AI's profile picture
狂热AI4 months ago

是因为算力充沛哈哈哈

Moussa . AI's profile picture
Moussa . AI4 months ago

Insane speed! 🔥 MiMo-V2.5-Pro UltraSpeed hitting 1000+ tokens/s on a full Pro model and dropping a complete interactive calculus web app in seconds? Xiaomi just crushed it. Respect 👏 I’ll trying right away..

Related Videos

给大家带来小米 MiMo-v2.6-pro 的测试速报! 简单来讲, 这次不愧是直播RL带来的效果, AgenticCoding 能力提升明显, 之前的 MiMo-v2.5-Pro 在我的向量数据库测试中得分只有2505, 而这次直接翻了3倍, 得分来到了7810. 与 Claude Fable-5 分数接近了. 而且算法也进化为了单图HNSW分层近邻图(M=16/M0=28) + AVX-512精确距离 + 每线程visited stamp. 可以说只要再优化一下算法细节就是SOTA水平了. 而前端测试则是MiMo需要提升的重点了, 这次演示效果虽然有提升, 但是无奈横评的其他模型都太强大了, 而小米的加分项也是本身的算法Coding能力, 比如它是可以 one-shot 写出一个光追引擎的. 但是由于空间理解/物理模拟能力不到位, 小球没办法完成把墙壁撞破的演示. 另外本次还给大家测试了输出速度, MiMo-v2.6-pro-ultraspeed 版本可以达到464tps的速度, 着实恐怖. 而根据MiMo的技术报告, 峰值可以达到900tps, 常态化请求也可以稳定在500tps左右, 考虑到模型 1.02T 的总参数量, 以及高达 42B 的激活参数量, 再加上同等规模模型一般都在60-80tps的水平, 这个成绩相当亮眼. 当然也有值得注意的点, 比如这次测试我发现它还是有早停的问题的, 上面的向量数据库测试, 每轮最大迭代50次, 但是三次测试中, 有两次它迭代到30轮左右就没办法提升分数, 于是选择了直接交卷, 浪费了测试机会. 以及, 模型思考偏长, 而且暂时不支持调整思考强度, 加上普速接口现在异常火爆, 所以最好给大一些超时, 避免无法输出结果. 总结: MiMo-v2.6-Pro 适合搞后端AgenticCoding开发! 要算法有算法要质量有质量. 但是前端空间理解和美学还需要加强. 考虑到直播后训练的时候看到的67%的训练语料都是Coding, 这个结果丝毫不惊讶了. 希望下个版本前端能力也有提升! #mimo #mimov26 #mimov26pro #x小米mimo

karminski-牙医

12,931 views • 1 month ago