Vor Luo Fuli's Debüt hat Xiaomi plötzlich etwas veröffentlicht. Der Code ist weltweit der stärkste und insgesamt mit DeepSeek-V3.2 vergleichbar. [Mit praktischer Prüfung]

According to a report by Zhidongxi on December 17th, today, Xiaomi released and open - sourced its latest Mixture of Experts (MoE) large - scale model MiMo - V2 - Flash, which has a total of 309 billion parameters and 15 billion active parameters. This morning, at the Xiaomi 2025 Xiaomi Smart Mobility, Home and Ecosystem Partners Conference, Ms. Luo Fuli, the person in charge of the Xiaomi MiMO large - scale model, will make her debut and deliver a keynote speech.

This model is specifically built for inference, coding, and Agent scenarios. It supports a hybrid thinking mode, allowing users to switch between the model "thinking" or giving an immediate answer. It can generate a fully functional HTML webpage with a single click and collaborate with coding frameworks such as Claude Code, Cursor, and Cline. The model provides a 256k context window and can complete tasks involving hundreds of rounds of Agent interaction and tool calls.

Benchmark test results show that the performance of MiMo - V2 - Flash is basically on par with DeepSeek - V3.2. It is only slightly inferior to DeepSeek - V3.2 in the "Last Human Exam" without any tool assistance and the creative text generation evaluation ARENA - HARD, but it has a lower latency.

In multiple Agent evaluation benchmarks, MiMo - V2 - Flash ranks among the top 2 globally in open - source models. In the code ability evaluation, it outperforms all open - source models and is comparable to the benchmark closed - source model Claude 4.5 Sonnet. However, its inference cost is only 2.5% of Claude 4.5 Sonnet's, and its generation speed is twice as fast.

MiMo - V2 - Flash can perform fast inference at a speed of 150 tokens per second. In terms of price, it costs 0.7 yuan per million input tokens and 2.1 yuan per million output tokens.

Xiaomi has open - sourced all the content of MiMo - V2 - Flash and contributed all the inference codes to the developer community SGLang. The API is free for a limited time. Currently, users can experience MiMo - V2 - Flash for free on the web.

Open - source address:

https://t.co/4Etm0yZKTL

Experience address:

https://aistudio.xiaomimimo.com

Technical report:

https://github.com/XiaomiMiMo/MiMo - V2 - Flash/blob/main/paper.pdf

01. Capable of writing long texts and creating beautiful webpages, but the actual test is not yet stable

Zhidongxi first tested MiMo's web development capabilities.

First, Zhidongxi asked it to generate an e - commerce page, which needed to include elements such as a product main image carousel, specification selection, price, and inventory. After the code was generated, MiMo would display a preview interface on the right. Except for the function of viewing large product images, which needs to be used in the full - version, other elements are almost the same as a real online shopping experience.

Then, to increase the difficulty, Zhidongxi asked it to generate a web - based maze game, which needed to have four difficulty levels and be controllable via mouse and keyboard. In the generated webpage, MiMo also added functions such as hints, scoring, and timing, making the entire interactive experience more complete.

Zhidongxi also asked it to generate an interactive webpage for "Christmas tree decoration". On the left side of the webpage, there are colored lights, bells, stars, gift boxes, and bows. Users can drag these items to decorate the Christmas tree on the right side. They can also clear all decorations with a single click or directly generate random decorations. Snowflakes will continuously fall on the entire screen, and background music can be played.

Regarding its chatting ability, Zhidongxi first asked MiMo to "comfort a lovelorn cat in a programmer's tone". MiMo told the kitten to "clear the cache" to forget unhappy things and also advised the kitten to "upgrade the hardware", which incidentally promoted Xiaomi.

Finally, Zhidongxi asked MiMo to generate a script for a science - fiction and suspense AI short drama. Right from the start, MiMo presented two key elements: a deep - diving memory extraction helmet and a Truth Ring AI ethics firewall. Through retrieval, it was found that MiMo incorporated these two objects into the subsequent plot, which directly served as the key to driving the plot forward.

In addition, Xiaomi also released many cases on its official blog.

First, in terms of webpage generation ability, its first official case was to "generate a beautiful and usable macOS simulation operating system using HTML". In the final generated interface, it included basic macOS - style desktop icons, and users could switch the background and open the Finder.

Secondly, it was required to create a high - fidelity 3D web application for a "luxurious interactive Christmas tree" with a luxurious visual effect. The result generated by MiMo can switch between the Christmas tree and a chaotic state and support gesture interaction.

The third demonstration was to create an interactive solar system explorer using HTML. When users click on each sphere, a brief introduction of the planet will appear.

Zhidongxi also asked MiMo to "create an educational interactive solar system explorer using HTML", but the final result was not as good as the official demonstration. When it was generated for the first time, the planets could not be selected. After regenerating, the options under the selectable planets simply disappeared.

In terms of chatting ability, MiMo can answer hard - core science - popularization questions such as "Do large - language models have knowledge?"

MiMo can also answer questions like "Should one take an immortality pill?" and will analyze from a human perspective whether to take it and what possible conditions there might be.

In addition, this model also has the ability to write long texts. It can "write a melancholy love story about a mis - delivered letter".

MiMo can also write profound non - fictional works. For example, it can center around a lonely elderly person and the emails he receives within a week to explore loneliness in modern society.

02. Specifically built for inference, coding, and Agent, comparable to K2 Thinking and DeepSeek V3.2 Thinking

In the 2025 American Invitational Mathematics Examination (AIME) and the scientific knowledge benchmark GPQA - Diamond, MiMo - V2 - Flash ranks among the top two in open - source models. In the Software Engineering (SWE) verification and multi - language benchmark tests, MiMo - V2 - Flash ranks first among all open - source models and is on par with the world's top closed - source models.

Judging from the overall benchmark test results, MiMo - V2 - Flash's performance in most inference benchmark tests is comparable to that of K2 Thinking and DeepSeek V3.2 Thinking, while maintaining competitiveness in high - quality open - ended answers.

In the long - context evaluation, MiMo's model outperforms K2 Thinking.

In Agent tasks, MiMo - V2 - Flash surpasses all open - source competitors in the SWE - Bench Verified and its performance is close to GPT - 5 - High. In the multi - language version of SWE - Bench, MiMo - V2 - Flash solved 71.7% of the problems.

In the search Agent evaluation, MiMo - V2 - Flash scored 45.4 on BrowseComp, which can be further improved to 58.3 with context management.

Xiaomi has shared all the inference codes with SGLang and open - sourced them. The results of community - tested single - machine performance are as follows:

Under the condition that the single - machine throughput of Prefill is about 50,000 tokens per second, it has achieved excellent Time to First Token (TTFT) performance for different context lengths.