Chinese company Moonshot AI has introduced its new AI model Kimi K3. The company claims that this is the world’s first open model with 2.8 trillion parameters.
It has a context window of 1 million tokens. Its scaling efficiency is about 2.5 times better than the previous model KM K2.
Let us understand this with a simple example: Suppose the AI model is like a human brain, which has to prepare for a huge exam.
1. What are the parameters?
You can think of parameters as the ‘experience’ of the AI. When a child grows up, the more things he learns, the faster his mind works. In the world of AI, the more parameters a model has, the more intelligent it is considered to be. Having 28 lakh crore parameters means that the brain of this model is extremely huge and it can solve very complex and difficult questions by understanding them like humans.
2. What is 3T class?
Here ‘T’ means trillion. Since this model has 2.8 trillion parameters, which is very close to 3 trillion (3.0T), it has been called a ‘3T class’ model. Reaching the 3T class in the world of AI is a huge achievement, because till now such big models were available only to select companies like America’s Open AI.
3. What is the context window of 1 million tokens?
context Window refers to the AI’s ‘memory’ or ability to read at once. When you chat with an AI, the length of a paragraph or document it can remember at a time is its context window.
Example: If you give 5-10 pages to a typical AI model, So he will read it, but if he gives the whole book he will feel exhausted. The Kimi K3’s capacity of 1 million tokens means you could give it an entire thick book, thousands of lines of computer code or a year’s worth of bank statements all at once; It will keep all that information in mind and give you an accurate analysis in a second.
4. What does it mean to have 2.5 times better scaling efficiency?
‘Scaling efficiency 2.5 times better’ means that the Moonshot company has made its software so efficient that it can do much bigger work, faster and with two and a half times less computer power and cost than its previous model.
5. What is the advantage of being an ‘open model’?
Till now American companies keep their big models hidden. This is called closed model. They just let you use it, but don’t tell you how it is made from inside. The Chinese company has kept it ‘open model’.
This means that she will release the ‘blueprints’ of this powerful brain for free or openly to software developers and scientists around the world. With this, any common engineer in the world can take this AI and change it as per his wish, create a new app or use it in his work.
Moonshot will open its secret formula on July 27
Moonshot AI blogged that Kimi K3 is built on ‘Kimi Delta Attention’ and ‘Attention Residuals’. It also has native vision capabilities. This model is designed for complex tasks like coding, knowledge work and reasoning. According to the company, its full model weights will be released by July 27.
1. What is ‘Km Delta Attention’ and ‘Attention Residuals’?
Understand with example: When you read a very thick book, do you pay equal attention to each word? No. Your brain pays ‘attention’ only to the main words and important lines and skips the unnecessary words.
Benefit: The new formula created by Moonshot AI named ‘Km Delta Attention’ and ‘Attention Residuals’ gives this super power to this AI. This technology does not let the AI’s mind wander when you give it a huge task or a file of millions of words.
It focuses only on the most important points and retains the ‘residuals’ of previous studies in the mind. Due to this, AI works very fast without getting tired and without making mistakes.
2. What are ‘Native Vision Capabilities’?
Understand with example: Earlier AI could only understand text. If you showed them a photo, they had to first convert that photo into words, only then they could understand. But ‘Native Vision’ means that the Kimi K3 has the ability to see.
Benefit: If you show it a photo, a complex graph, a medical scan (like X-ray or MRI) or a video, it will immediately understand it by looking at it like a human eye and answer your questions on it.
3. What is meant by coding, knowledge work and reasoning?
The company says that this model will do three big things:
Long-Horizon Coding: This does not mean just writing small codes. Like software engineers, he can write a large computer program or app consisting of thousands of lines simultaneously and find and fix the shortcomings in it.
Knowledge Work: He can work like researchers, scientists and big lawyers. Searching through years old files, analyzing big legal cases or preparing a complete research report on any topic is a snap.
Reasoning: It means ‘ability to think, understand and reason’. It will not just tell you rote things, but like humans, will help in taking the right decision by applying ‘ifs and buts’ in difficult mathematical problems or business problems.
Kimi K3 lags behind America’s latest AI models
The company admitted that the performance of Kimi K3 lags behind American models like Claude Fable 5 and GPT 5.6 Sol. However, in many benchmark tests it has left behind the older American models. Moonshot AI is supported by Chinese companies Alibaba and Tencent.
Chinese firms stole data by creating 24 thousand fake accounts
Meanwhile, American companies have expressed concern over Chinese firms releasing models cheaply. Anthropic has alleged that DeepSeek, Moonshot, and MiniMax conducted large-scale campaigns to illegally extract the capabilities of their models. According to Anthropic, the cloud was used 16 million times from about 24,000 fake accounts,