What Is Kimi and How Does It Work? Moonshot AI's Chatbot and K3 Model Explained

Kimi is the AI assistant app from Moonshot AI, a Beijing-based company that's been making serious waves since mid-2026. If you've heard the name floating around but aren't quite sure what it is, you're not alone. It went from a China-focused assistant to a globally recognized name pretty fast, and a lot of that has to do with one thing: its K3 model.

Released in July 2026, Kimi K3 is a 2.8-trillion-parameter model. That's the largest open-weight AI model in the world right now, and it's completely free to download, modify, and use. Moonshot AI didn't just build a chatbot. They built one of the most capable open-source AI systems ever released, and then gave it away.

One of the most common questions about Moonshot AI's Kimi is which country it comes from. China. We'll get into the details below.

This article covers what this AI assistant actually does, who's behind it, how the K3 model works, what it costs, and everything else worth knowing before you try it.

1. Who Made Kimi and Where Does It Come From?

Kimi comes from Moonshot AI, a Chinese artificial intelligence company founded in April 2023 and headquartered in Beijing. The company's Chinese name translates literally to "the dark side of the moon." That's not an accident. Founder Yang Zhilin and his co-founders named it after Pink Floyd's 1973 album The Dark Side of the Moon. They're all rock music fans. Yang himself played drums in a college band called Splay.

Yang Zhilin is worth knowing about because his background explains a lot about why the company moves the way it does. He graduated top of his class from Tsinghua University's computer science program, then got his PhD from Carnegie Mellon in four years instead of the usual six. He co-authored two of the most influential papers in natural language processing: Transformer-XL and XLNet, which together have over 22,000 academic citations. Before starting Moonshot AI, he worked at Google Brain and Meta AI.

The company raised money at a pace that's hard to ignore. In the first half of 2026 alone, Moonshot AI pulled in roughly $60 billion across multiple funding rounds. Backers include Alibaba, Tencent, Meituan, and several major venture capital firms. As of mid-2026, its valuation sits north of $300 billion. The company is actively preparing for an IPO on the Hong Kong Stock Exchange, expected by early 2027.

So: Chinese company, Beijing headquarters, founded 2023, led by a Carnegie Mellon PhD who wrote foundational AI research before most of the current AI boom even started. That's the short version.

2. What Is the Kimi K3 Model?

Kimi K3 is the model that runs underneath the assistant. It's what gives Kimi its capabilities, and it's also available as a standalone system developers can use through an API.

The numbers are big. 2.8 trillion parameters total, structured as a Mixture-of-Experts (MoE) model with 896 expert modules. For each token it processes, only 16 of those experts activate. That's the trick that makes a model this large practical to run. You get the knowledge and reasoning capacity of a 2.8-trillion-parameter system while only paying for the compute of a much smaller active footprint, around 104 billion parameters per inference step.

It supports a 1-million-token context window. In practical terms, that means you can feed it the entire text of War and Peace and still have room for questions. Or drop in a full codebase, a stack of legal documents, or every research paper on a topic and ask it to synthesize across all of them at once.

K3 is a major architectural departure from the previous K2 model. It introduces something called Kimi Delta Attention (KDA), a hybrid linear attention mechanism designed specifically for processing extremely long sequences without the usual slowdown and quality loss that hits most models at long contexts. There's also a new visual encoder called MoonViT-V2 that was trained from scratch, giving K3 native image and video understanding without needing external vision plugins. Moonshot AI says the overall scaling efficiency is roughly 2.5 times better than K2.

On benchmarks, the results are genuinely competitive with the best closed-source models. K3 topped the Arena.AI Frontend Code Arena leaderboard with 1,679 points, ahead of Claude Fable 5 and GPT-5.6 Sol. On the Artificial Analysis Intelligence Index, it ranked third globally with a score of 57, behind only the top two closed-source models. On the AA-Briefcase knowledge work benchmark, a test simulating real office tasks with thousands of emails, Slack messages, and financial documents, K3 scored 1,543 points, beating GPT-5.6 Sol's 1,501.

None of these numbers mean it's "better" in every way. Closed-source models still lead in certain reasoning and general-purpose tasks. But for coding, long-context work, and agent-based task execution, the K3 model is in the top tier. And all of it is open-source. Moonshot AI released the full model weights on July 27, 2026, under a Modified-MIT license. Within 72 hours, the model hit over 180,000 downloads on Hugging Face and spawned more than 340 community fine-tuned variants.

3. What Can You Actually Do With Kimi?

Kimi as an app is available on Android, iOS, HarmonyOS, the web, and through a desktop client called Kimi Work. Here's what it covers.

Long-document analysis. This is what built Kimi's reputation as a long-text AI. You upload a PDF, a research paper, a contract, or a book, and it reads the whole thing. Then it answers questions, summarizes sections, or compares documents. Lawyers use it to review contracts. Researchers use it to synthesize literature across dozens of papers at once. For anyone who works with long texts, this is the core use case.

Coding. Kimi Code and the K3 API are built for software development work. The model can handle multi-file, multi-step projects. It's not just generating snippets. It reads codebases, understands architecture, debugs, refactors, and writes tests. In one documented case, K3 autonomously optimized a financial trading engine over a 13-hour session, iterating through 12 strategies and modifying thousands of lines of code to deliver a 185% throughput improvement. Developers have also used it to build full interactive web apps, 3D visualizations, and working system emulators from a single prompt.

Research and writing. It searches the web, pulls in information from multiple sources, and produces structured reports, articles, or presentations. The K3 Cluster mode is particularly useful here. It spins up multiple agent instances simultaneously, each working on a different piece of a large task, then combines the results. In the API's first 23 days after the K3 launch, it handled over 17.3 billion calls.

Multimodal interaction. K3 understands images and video natively. You can show it a screenshot of a UI and ask it to write the code to recreate it. Or feed it a video and ask for a summary with timestamps. This isn't a separate feature bolted on. The vision understanding is built into the model itself.

Agent-based automation. Kimi supports goal-based automation where you describe a task and it figures out the steps, calls tools, writes and runs code, and delivers a finished result. The Agent Swarm feature coordinates multiple agents working in parallel on different parts of a complex project. Elon Musk called the K3 release "impressive" on social media. Fair enough.

For everyday users, the free tier is generous enough for casual use. Power users who need the full 1-million-token context or long-running agent sessions will want the paid plan.

4. Kimi Pricing

Kimi has a few different pricing tiers depending on how you use it.

The free tier gives you access to the basic K3 model with standard context length. It's not time-limited or trial-based. You can use it indefinitely for regular tasks like chatting, document analysis of reasonably sized files, and web search.

The paid subscription is called Allegro, with three levels: 49 RMB, 99 RMB, and 199 RMB per month (roughly $7, $14, and $28 USD, with exact amounts depending on exchange rates). The higher tiers give you the full 1-million-token context window, priority access during high-traffic periods, and the K3 Cluster agent swarm feature. They also come with higher usage quotas. The mid and top tiers are aimed at professionals who use Kimi as a daily work tool.

For developers, the API is pay-per-token with pricing that's significantly cheaper than comparable closed-source models. Cache hits bring the input cost down significantly. Moonshot AI's strategy here is clear: make it cheap enough that switching from a more expensive AI provider is a no-brainer.

One thing to know: after the K3 release in July 2026, the surge in demand was so intense that Moonshot AI temporarily paused new user subscriptions on July 20 to manage server capacity. It's a sign of how fast this AI tool took off, but it also means availability can be tight during peak periods. The company has since expanded its compute infrastructure and the situation has improved.

5. Kimi Work: The Desktop Agent

In June 2026, Moonshot AI launched Kimi Work, a desktop application that runs as a local AI agent on your computer. The short description: it's the version of Kimi that actually does stuff instead of just talking about it.

It runs on Mac and Windows. Once installed, it can read files on your computer, operate your browser, run code, and generate documents. You describe what you want, and it figures out the steps.

The standout feature is the ability to spawn up to 300 sub-agents that work in parallel. Give it a task like "research the last ten years of Warren Buffett's portfolio and create a presentation," and it splits the work: some agents pull data from financial websites, others analyze your local files, others format slides. The results merge into a finished deliverable.

It also operates your actual browser, not a sandboxed one. That means it can work with sites you're logged into, use your saved credentials, and access subscription-based data platforms. For knowledge workers in finance, law, and research, this makes a real difference.

The desktop client itself was built with heavy AI assistance. According to Moonshot AI, about 92% of the code in the initial Kimi Work release was generated by their own coding agents, with human engineers supervising and integrating the output. Over 50,000 lines of effective code were produced in about a week.

Kimi Work is currently in public beta. It runs on the K3 model and is free to use during the beta period. Some advanced features require an Allegro subscription.

6. Kimi Privacy and Data Handling

This gets complicated. And it matters if you're planning to use Kimi for anything sensitive.

Moonshot AI is a Chinese company, which means it operates under Chinese data regulations. The company collects conversation data to improve its models, as most AI companies do. If you're using the free consumer version, your data may be used for training purposes. The paid enterprise tier and API plans have separate data handling agreements with more restrictive terms.

The open-source K3 model changes the privacy equation entirely. Since the weights are fully available, you can run K3 on your own hardware with no data leaving your network. Several cloud providers, including Alibaba Cloud, Tencent Cloud, and Huawei Cloud, offer one-click deployment options. For organizations with strict data sovereignty requirements, this is the route to take.

If privacy is a primary concern, the self-hosted option basically eliminates the issue. You control the data because you control the hardware. No gray area there.

7. What's Next for Kimi

Moonshot AI isn't slowing down. At the K3 launch celebration in Beijing, the team previewed K4, internally codenamed "Tidal," already in early verification on thousand-GPU clusters. Details are thin, but the direction points to an even larger model with deeper integration into real-world applications, possibly including satellite data processing and aerospace partnerships.

The company is also preparing to go public on the Hong Kong Stock Exchange, likely within the next six months. If the IPO happens as expected in late 2026 or early 2027, it would be one of the largest AI listings to date from a Chinese company.

On the product side, the Kimi Work desktop agent is still in public beta, but it's being positioned as a core part of the company's strategy. The vision is straightforward: move from a chatbot people talk to, to an agent that gets work done.

At the same time, K3's open-source release has triggered what some analysts are calling a "second DeepSeek moment." When a model this capable goes fully open, it reshapes the competitive field. US companies like Microsoft have already evaluated integrating K3 into Azure and Copilot. The downstream effects on API pricing, developer tooling, and enterprise adoption are still playing out.

Conclusion

Kimi is a Chinese AI assistant powered by one of the most capable open-source models available today. It's built by Moonshot AI, a Beijing company founded in 2023 that has since grown into one of the most valuable AI startups in the world. The K3 model puts it in the same conversation as GPT-5.6 and Claude Fable 5, particularly for coding, long-document work, and agent-based tasks.

It's free to try, openly available for developers to build on, and backed by a company that's investing aggressively in model research. The Kimi Work desktop agent adds a practical layer for people who want AI to actually do things rather than just answer questions.

If you want to try Kimi for yourself, you can download the Android app right here. It gives you access to the full K3 model on your phone, with document upload, web search, and multimodal input all built in.