A Chinese company called Moonshot AI just made history. On July 16, 2026, they released Kimi K3, the biggest open AI model the world has ever seen. It has 2.8 trillion "parameters" — the building blocks that make an AI model smart.
To put that in perspective, no open AI model has ever reached this size before. Kimi K3 entered what experts call the "3T class," a tier no openly available model had reached until now.
But the size is not the most interesting part. What is really surprising is how Moonshot built it — and what it means for the global AI race between China and the United States.
What Is an Open AI Model?
Before we dive in, let us explain "open." Most big AI models, like ChatGPT, are closed. You can use them, but you cannot see inside or run them yourself. The company owns everything.
An open-weight model (sometimes called open-source) is different. The company shares the actual "brain" of the AI. Anyone can download it, study it, modify it, and run it on their own computers. It is like getting the recipe for a secret sauce instead of just being served the dish.
Kimi K3 is open-weight. That means companies, researchers, and even governments can run it themselves.
Why Kimi K3 Is Special
Most AI models need two things to work: compute (the brain power to calculate answers) and memory (storage to hold the model ready while it works).
The United States has strict rules that limit China's access to powerful AI chips — the hardware that provides compute. So Moonshot made a clever move. Instead of trying to get more compute, they designed Kimi K3 to use less compute by relying more on memory.
Here is how they did it:
- Mixture-of-Experts: Instead of using the whole model for every word, Kimi K3 splits into 896 specialist sections. For each word, it only activates 16 of them — about 1.8%. This slashes the compute needed.
- Smaller Memory Footprint: They trained the model to work at 4 bits of precision instead of the usual 16. This shrinks the model from 5.6 terabytes down to about 1.4 terabytes. That is a huge savings.
- Delta Attention: A new trick that makes the model work faster when reading very long documents (up to a million words, or a few thousand pages).
How Good Is It?
Kimi K3 is impressive, but it is not perfect. Here is the honest scorecard:
- In blind tests by developers, Kimi K3 ranked first for writing frontend code, beating out major competitors.
- However, Moonshot itself admits that overall, Kimi K3 still trails behind top models like Claude Fable 5 and GPT 5.6.
- The model can sometimes make unexpected decisions when it is unsure what the user wants.
- Its generation quality can become unstable in certain situations.
So it is a strong model, especially for coding, but not the absolute best at everything yet.
What Does It Cost?
Here is where it gets interesting for businesses:
- $3 per million input tokens (the words you send to the model)
- $15 per million output tokens (the words the model generates back)
- Prices drop to $0.30 for input when the model has recently seen similar content
That is much cheaper than some competitors (like Fable 5 at $50 for output), but more expensive than others (like DeepSeek V4 at $0.87 for output). Kimi K3 is no longer the budget option it used to be.
Why This Matters to Regular People
You might be thinking, "Why should I care about a Chinese AI model?" Here are three reasons:
- More Competition Means Better AI: When China and the US compete, everyone wins. Prices drop, models get smarter, and new features arrive faster.
- Open Models Give You Control: Open models let companies and governments run AI on their own servers. Your data never leaves your control. This matters for banks, hospitals, and anyone who cares about privacy.
- The AI Race Is Global: The US chip restrictions were meant to slow China down. Kimi K3 shows that creative engineering can find workarounds. The gap between the two countries may be closing faster than expected.
The Big Picture
Kimi K3 proves that the AI race is far from over. Even with hardware restrictions, Chinese companies are finding clever ways to build world-class AI. The open-weight movement is growing fast — open models now handle 29% of all AI requests through one major service, up from about 11% just three months ago.
The full Kimi K3 model will be available for download on July 27, 2026. After that, we will finally see how well it performs in the real world.
One thing is clear: the future of AI is not just about who has the most powerful chips. It is also about who has the smartest ideas.
