Varto
Хочете будувати кар’єру в Data Science та AI? Vodafone Big Data Lab відкриває новий набір —>

Xiaomi released MiMo-V2.6-Pro, which topped the ranking of AI models with open parameters. According to the Artificial Analysis result, it was equal to Grok 4.7

Xiaomi has introduced a new generation of AI models MiMo-V2.6. The older MiMo-V2.6-Pro immediately topped the rating of models with open parameters from Artificial Analysis. It scored 46 points. Xiaomi positions MiMo-V2.6 as the basis for AI agents that can work on a task for a long time, use tools, write code and correct their own mistakes.

Leave a comment
Xiaomi released MiMo-V2.6-Pro, which topped the ranking of AI models with open parameters. According to the Artificial Analysis result, it was equal to Grok 4.7

Xiaomi has introduced a new generation of AI models MiMo-V2.6. The older MiMo-V2.6-Pro immediately topped the rating of models with open parameters from Artificial Analysis. It scored 46 points. Xiaomi positions MiMo-V2.6 as the basis for AI agents that can work on a task for a long time, use tools, write code and correct their own mistakes.

This is reported by VentureBeat.

A model with open parameters means that Xiaomi has published its trained parameters. So developers can download the model and run it themselves. MiMo-V2.6-Pro is available on Hugging Face. It can also be retrained for your own tasks. Xiaomi has released the model under the MIT license, which also allows commercial use.

MiMo-V2.6-Pro has about 1 trillion parameters, although only 42 billion are activated during operation. It can also work with images, audio and video. The context window reaches 1 million tokens. This is enough for the model to keep in memory a very large array of documents, code or a long history of the work of an AI agent.

Along with the Pro, Xiaomi released a "younger" model, MiMo-V2.6-Flash. For the Pro, a million incoming tokens via the API cost $0.435, outgoing tokens cost $0.87. For Flash, they cost $0.14 and $0.28, respectively.

Xiaomi also shared how the model training process went. During the reinforcement learning stage, the AI ​​performed tasks many times, received feedback on the results, and gradually learned to find better solutions. Xiaomi conducted 30 large cycles of such training for the Pro and Flash — a total of about 750,000 attempts to complete various tasks.

But at some point, the models learned not only to solve problems, but also to find ways around them. For example, instead of fixing a bug in the code itself, the AI ​​could download a newer version of the package, find a ready-made file with a fix, or look in the history of the repository where the desired solution already existed. So they were, in effect, “spying” on solutions.

So Xiaomi began to remove such “loopholes”. Later versions of the code, which could already contain fixes, as well as caches and other files that could give away the answer, were removed from the training environment. The company even created a separate “hack agent” that, before training, specifically looked for ways to bypass the task and deceive the evaluation system. So, according to the company, in the final training, the share of confirmed cases of such “eavesdropping” remained below 2%.

Previously, dev.ua wrote that almost simultaneously with MiMo-V2.6, SpaceXAI released Grok 4.7 for programming and complex work tasks.

China's Moonshot unveils world's largest open-source AI model
China's Moonshot unveils world's largest open-source AI model
On the topic
China's Moonshot unveils world's largest open-source AI model
AI startup of former OpenAI — Thinking Machines — launched an artificial intelligence model with open weights for 975 billion parameters
AI startup of former OpenAI — Thinking Machines — launched an artificial intelligence model with open weights for 975 billion parameters
On the topic
AI startup of former OpenAI — Thinking Machines — launched an artificial intelligence model with open weights for 975 billion parameters
Read the country's main IT news in our Telegram
Read the country's main IT news in our Telegram
On the topic
Read the country's main IT news in our Telegram
Also Read
There's nothing to breathe. How to determine the level of air pollution and what to do when the city is drowning in smoke after shelling
There's nothing to breathe. How to determine the level of air pollution and what to do when the city is drowning in smoke after shelling
There's nothing to breathe. How to determine the level of air pollution and what to do when the city is drowning in smoke after shelling
Kyiv residents are once again experiencing air quality problems in the Ukrainian capital: visibility has significantly deteriorated, a strong smell of smoke has appeared in certain areas, and breathing has become difficult. The reason for this phenomenon is another Russian shelling, which has caused numerous fires in various areas of the capital and in some places has led to pollution levels exceeding the norm by dozens of times. We remind you of what you need to know about air quality and what useful tools are available for monitoring it.
Штучний інтелект DALL-E навчився домальовувати картини. Як це виглядає
Штучний інтелект DALL-E навчився домальовувати картини. Як це виглядає
Штучний інтелект DALL-E навчився домальовувати картини. Як це виглядає
2 comments
Штучний інтелект почав озвучувати фільми на MEGOGO
Штучний інтелект почав озвучувати фільми на MEGOGO
Штучний інтелект почав озвучувати фільми на MEGOGO
6
Штучний інтелект навчився реставрувати старі фотографії, перетворюючи їх на якісні зображення: відео
Штучний інтелект навчився реставрувати старі фотографії, перетворюючи їх на якісні зображення: відео
Штучний інтелект навчився реставрувати старі фотографії, перетворюючи їх на якісні зображення: відео
2 comments

Have important news to share? Message our Telegram bot

Key events and useful links in our Telegram channel

Discussion
No comments yet.