China’s New Kimi K3 Model Has the American AI Industry on High Alert

Moonshot AI has released Kimi K3, an artificial intelligence model that excels at coding and agentic tasks. Here’s what we know about the largest open-weight model in the world so far.

Written by Matthew Urwin
Published on Jul. 28, 2026
The Kimi logo displayed on a black phone screen, with a black background illuminated by a purple light.
Image: Shutterstock
REVIEWED BY
Summary: Designed by Chinese startup Moonshot AI, Kimi K3 boasts nearly 3 trillion parameters, making it the largest open-weight model to date. We’re exploring the model’s architecture, capabilities and performance compared to its American counterparts — and whether it’s time for U.S. leaders to panic.

The artificial intelligence industry may have just witnessed its next DeepSeek moment courtesy of Moonshot AI. On July 16, the Chinese startup released Kimi K3, an open-weight AI model boasting 2.8 trillion parameters and impressive coding abilities — all at a much lower price than models developed by the likes of OpenAI and Anthropic. In fact, it’s already stoking fears that the United States has officially lost its lead in the AI race against China. 

What to Know About Kimi K3

Built by Chinese startup Moonshot AI, Kimi K3 is an AI model that excels at coding, knowledge work and video and game development. It contains 2.8 trillion parameters and a context window of 1 million tokens, making it the largest open-weight model in the world. K3 represents a major stride forward for Chinese AI, closing the gap with the United States and signaling a more heated chapter of the global AI race.

But is it really time for U.S. officials and AI companies to start panicking? We’re taking a closer look at how Kimi K3 works, how it stacks up against its American counterparts and how it could ultimately influence the Trump administration’s AI policies as the country struggles to maintain its technological edge over China. 

Another Aspect of the AI RaceThe AI Chip Race Is Heating Up. Will Anyone Catch Nvidia?

 

What Is Kimi K3? 

Kimi K3 is a massive large language model, with nearly 3 trillion parameters and a context window of 1 million tokens. Its sheer size enables it to excel at advanced coding, reasoning and knowledge work. Meanwhile, built-in visual understanding helps it handle tasks in areas like game development and video editing

The model is open-weight, which is not the same as open source. Rather than disclose all the details of Kimi K3, Moonshot AI revealed only the parameters of the model’s final version post-training. Users can freely access these parameters to get a sense of how K3 works, run it on their own machines and tailor it accordingly. At the same time, though, they can’t see the training data or training code, preventing them from fully replicating the model. 

 

How Does K3 Work? 

Kimi K3 relies on several mechanisms that form an intricate neural network system to power the model’s advanced reasoning

  • Kimi Delta Attention (KDA): A linear attention mechanism that updates a continuous running memory with each new token. This approach lets K3 more selectively decide which information to remember or forget. 
  • Attention Residuals (AttnRes): A neural network architecture that allows each layer to recall only relevant outputs from previous layers, instead of compiling all outputs. This method is designed to make K3’s memory more cost-efficient.  
  • Stable LatentMoE: A neural network architecture that breaks up K3 into various subnetworks, or “experts,” activating a mere 16 of the model’s 896 experts per token. This strategy improves K3’s efficiency. 

Together, these mechanisms enable Kimi K3 to complete more demanding tasks while keeping costs and compute down. 

Perhaps K3’s greatest feature, though, is how it uses its “thinking traces.” The model employs chain-of-thought, a technique that breaks down a problem into a series of steps to be solved sequentially. During this reasoning phase, previous models could only handle specific tasks or shorter workflows, like deciding on a webpage’s visual design, planning sections or writing code. K3 can cover all of these steps, essentially simulating how an AI agent operates within its chain-of-thought process. 

The result is that Kimi K3 spends more time and energy iterating on its work, producing better outputs at the expense of using up more tokens. It’s a trade-off that seems to be paying dividends for the model’s overall performance, particularly when it comes to coding.

More on U.S.-China RelationsA Guide to How America Controls Nvidia Chip Exports to China

 

What Can K3 Do? 

Besides coding, Kimi K3 excels in several other areas, including chip design, knowledge work, video editing and game development. 

Long-Horizon Coding 

K3 can sustain longer software engineering sessions, operating in the background while a user pivots to other projects. The model has the size to process large volumes of data, making it ideal for situations involving extensive code repositories. It can also orchestrate terminal tools without much human supervision

Chip Design 

According to Moonshot AI, Kimi K3 crafted a chip for a smaller AI model with the same architecture. It took just 48 hours for K3 to finish the chip, tapping open-source exploratory data analysis tools to inform its design. 

Knowledge Work 

Kimi K3 surpassed OpenAI’s GPT-5.5 and Anthropic’s Claude Opus 4.8 on Moonshot AI’s internal benchmarks for knowledge work, or tasks that involve synthesizing information, exercising judgement and coming up with outputs similar to those produced by many white-collar professionals. The model’s expertise can be applied to scientific research as well. For instance, K3 can be used to review published code, generate corrections and produce an analysis for further investigation. 

Video and Game Development 

Because of its multimodality, Kimi K3 can naturally adjust to any task involving visual creation. It works with text, images and video, making it suitable for video editing, animation and motion design. The model can also combine computer-aided design with its coding features to convert images and videos into full-fledged video games

 

How Does K3 Compare to American AI Models? 

Based on Moonshot AI’s testing, Kimi K3 held its own in coding and general agentic tasks against top models, including GPT-5.6 Sol, Fable 5 and GPT-5.5. It even consistently beat GPT-5.5 and Claude Opus 4.8 on several internal knowledge benchmarks. Such progress is particularly impressive when considering that DeepSeek’s V4 Pro model was estimated to be eight months behind similar models built by OpenAI and Anthropic, according to a 2026 study by the Center for AI Standards and Innovation (CAISI). 

Although K3 has quickly closed the gap between Chinese open-weights and American proprietary models, it still has some ground to make up. Hugging Face notes that the model doesn’t always ask for clarification when it should before acting, and its performance can be affected by agents that tweak its chain of thought. It doesn’t have the best user experience design, either, compared to Fable 5 and GPT-5.6 Sol. 

A joint study by CAISI and the UK Artificial Intelligence Security Institute (UK AISI) also found that Kimi K3 performs poorly on cybersecurity evaluations, with its safeguards failing to prevent it from attempting to exploit software systems. That said, Kimi K3 falling short of American frontier models has done little to ease the concerns of U.S. officials and tech leaders, who are already preparing a counterpunch.

Why American AI Supremacy Is UncertainEurope Pushes Back on Trump’s AI Plan

 

What Does Kimi K3 Mean for the U.S.-China AI Race? 

The U.S. response to Kimi K3 has been sour, to say the least. Shortly after K3’s release, President Trump’s top science advisor Michael Kratsios accused Moonshot AI of illegally distilling Anthropic’s Fable to train K3 — an unfounded claim — and U.S. Treasury Secretary Scott Bessent threatened to add it to the “Entity List,” which subjects organizations to trade restrictions. Meanwhile, Big Tech CEOs worry about competing against Chinese open-weight models in the American market and what that could mean for national security

Even so, American leaders may have no one to blame but themselves for the country’s dwindling lead over China in the AI race. After all, the Pentagon spent months fighting to designate Anthropic as a national security threat because it limited government access to Claude. And U.S. officials have blamed state regulations and data center moratoriums without acknowledging how vague federal frameworks have contributed to a confusing policy landscape that could further slow American innovation

In contrast to the government’s muddled response, American AI companies are moving quickly to reclaim their lead. For instance, Nvidia CEO Jensen Huang has pushed back against trade restrictions and AI-driven fears. Rather than cut off access to its models, Nvidia has formed a security alliance with U.S. and global partners to promote open-weight and open-source AI alongside AI safety. Anthropic has adapted as well by releasing a cheaper alternative to Claude Fable 5 called Claude Opus 5 in an effort to match Kimi K3’s cost-effectiveness. 

These measures are a promising sign for the American AI industry. But unless the U.S. government can unite with business leaders, local officials and everyday people on the direction of AI development, the country may continue to struggle to maintain its edge as Chinese tech companies ramp up the pressure on their American rivals.

Frequently Asked Questions

Kimi K3 is an open-weight AI model with 2.8 trillion parameters and a context window of 1 million tokens, enabling it to handle complex coding, reasoning and knowledge work. An intricate neural network system processes information efficiently, so K3 can handle massive computations without unnecessary energy or expenses. Chinese startup Moonshot AI created and announced the model on July 16.

No. Kimi K3 is an open-weight model, meaning that Moonshot AI only released its post-training parameters to the public. Users can gain a basic understanding of how K3 works, run it on their own computers and customize it for specific tasks. But they can’t fully reproduce the model since they can’t view its training code or data.

Because of its reasoning and multimodal capabilities, Kimi K3 excels at completing tasks in several areas: 

Moonshot AI found that Kimi K3 keeps pace with OpenAI’s and Anthropic’s top models on coding and general agentic benchmarks, and it consistently outperformed them on internal knowledge benchmarks. That said, K3 sometimes acts before asking for clarification, features a less-friendly UX design and can be influenced by agents that adjust its chain of thought, according to Hugging Face researchers. A joint study by two AI research organizations also discovered that K3’s safeguards don’t stop it from attempting to exploit security systems.

Many U.S. tech companies are worried about Chinese open-weight models entering the American market and have called for measures to block them. Players like Anthropic are also building cheaper models to counter Kimi K3’s lower cost. At the same time, Nvidia has downplayed concerns, choosing instead to establish a security alliance that supports the safe use of open-weight and open-source models among U.S. and international partners.

Explore Job Matches.