Kimi K2: What's the Fuss About and How Do You Use It?

SECTION 1

Introduction

A new large language model, Kimi K2,is causing a stir in the tech industry. Developed by Chinese AI company Moonshot AI, backed by digital giant Alibaba, its release is being hailed as another DeepSeek moment. Like DeepSeek, Kimi K2 is open-weight, meaning its trained parameters are freely available to download and adapt. Also like DeepSeek, it has shown impressive performance against established models.

SECTION 2

Architecture and how it works: mixture-of-experts and an agentic approach

### A mixture-of-experts architecture built for efficiency

Kimi uses a mixture-of-experts model architecture. This means it's made up of a system of separate sub-networks, each specializing in distinct parts of a given problem, of which only a fraction is activated to handle each request. The advantage of this approach lies in speed and computational efficiency.

Despite its 32 billion active parameters (and a trillion total), it's relatively cheap to run. While Claude Opus 4 costs $15 per million input tokens and $75 per million output tokens, Kimi costs a fraction of that at $0.15 per million input tokens and $2.50 per million output tokens.

### An agentic approach rather than step-by-step reasoning

Kimi is also notably described as an agentic LLM. Moonshot AI states in its product documentation that it was "carefully optimized for agentic tasks." This sets it apart from the reasoning approach commonly used in other established models, where the model is built to follow a sophisticated step-by-step approach to problem-solving. By contrast, the agentic approach Moonshot emphasizes is meant to let the model "learn" from external experience, a point raised by researchers David Silver and Richard Sutton in their paper The Era of Experience (which the Moonshot AI team specifically cites).

SECTION 3

Is Kimi K2 Really Another DeepSeek Moment?

Kimi K2's release has been described by some as another DeepSeek moment. That's not hard to understand, once again, we have an open model built by a Chinese company apparently outperforming established players. While that's true, it hasn't had the same cultural and economic impact as the story from earlier in the year.

Perhaps, in itself, that matters, the fact that high-performing open-weight models coming out of China no longer surprise us is a signal that the AI market is maturing.

That doesn't mean companies like OpenAI and Anthropic won't continue to dominate, but the technical innovations we've seen from DeepSeek and Kimi represent the diversity of the AI field right now and will no doubt encourage further innovation and experimentation.

@cite:demystifier-deepseek

SECTION 4

Kimi K2's Coding Performance and Feedback from the Field

One area where Kimi K2 has shown promise is coding tasks. According to Moonshot, the model has outperformed other established models on several benchmarks. Of course, proof of its real-world coding effectiveness will come from real-world use, but feedback so far has been positive from many software developers who've tried it.

The fact that Kimi K2 can be integrated with Claude Code, Anthropic's agentic coding tool, means we'll likely hear a lot more about Kimi K2's performance in the coming weeks.

To get a firsthand perspective on Kimi K2, Adservio software engineer Zhenjia Zhou kindly answered a few questions. He's been experimenting with the model in personal projects since its launch.

### Why he adopted it and what he likes about it

Richard Gall: Why did you start using Kimi K2? What drew you to it?

Zhenjia Zhou: I used Kimi K2 on the day it launched! I find Claude Sonnet 4 too expensive, especially for personal projects. So I found a way to use it with Claude Code, mainly for backend Python code.

RG: Are there any notable differences compared with other models?

ZZ: When I use Cursor, I usually use OpenAI o1. Compared with o1, Kimi is smarter when it comes to tool calling. For example, I like using the Sequential Thinking MCP server, which o1 doesn't really like to call. Most of the time it only calls it if I specifically ask for it in my prompt: in other words, I have to write "Please use sequential thinking to solve this problem." Claude Sonnet 3.7 has similar issues.

RG: What do you like about it?

ZZ: It's cheap and open source! Claude Sonnet 4 is quite expensive. For example, a task with Sonnet 4 can cost me between $10 and $20. But with Kimi K2, I can do about ten similar tasks for just 50 RMB ($7 USD). And since it's open source, I could potentially deploy it myself, which would make it even cheaper.

That means I can be a lot more efficient and productive. I can work on several tasks in parallel, if the ten tasks I'm working on don't conflict, I can just open ten separate instances of Claude Code and use Kimi K2 to work on each task.

I pictured working like this back when Claude Code first launched, but if you use Claude Sonnet 4, you'd spend a lot of money.

### Limitations observed and where to draw the line on use cases

RG: Are there any challenges or things you don't like about Kimi K2?

ZZ: I've found Kimi K2 quite slow. Right now, it takes longer to generate a response compared with Sonnet 4. I also think the context window is fairly small compared with Sonnet 4.

RG: When do you think you might use it over other more established models?

ZZ: Based on my experience, I don't think Claude Code is the best tool for Kimi K2, even though it's cheaper to use Kimi. Claude Code was, after all, designed for Claude Sonnet 4,when I use Kimi K2 with Claude Code, it's like there's a different soul in Claude Code's body!

That said, if Kimi K2 eventually gets a better interface than Claude Code, maybe I'd start using it instead of Claude.

RG: I've seen this called another DeepSeek moment, is it really?

ZZ: I think it shows that open source language models can play an important role in the AI landscape, not just in terms of cost but also in terms of performance.

RG: What do you like about open models?

ZZ: I think there are two appealing things about open models. One is that for companies that really care about privacy, they can deploy the model themselves. The other is that openness means more providers. Right now, for example, Claude Sonnet 4 is only available on AWS and Claude. That means they control the API price.

For open source models, there will inevitably be more providers on the market and maybe even price wars, which could lead to cheaper API prices.

Thanks to Zhenjia for taking the time to talk. It's certainly very early days for Kimi K2 adoption, we'll be watching it closely and perhaps even running our own experiments with it in the months ahead.

Disclaimer: The statements and opinions expressed in this article are those of the author(s) and do not necessarily reflect Adservio's positions.

SECTION 5

How to Seriously Evaluate Performance Claims

Figures published by a model vendor, whether Moonshot, Anthropic, OpenAI, or Google DeepMind, always deserve a degree of methodological caution. A public benchmark such as SWE-bench, LiveCodeBench, or Humanity's Last Exam measures a specific skill under specific conditions; it doesn't guarantee a direct transfer to an organization's real-world use cases, with its own data, its own latency constraints, and its own legacy code.

That's why building your own evaluation set, representative of the tasks teams actually face, remains the most reliable approach before choosing between models or generalizing a use pattern like the one Zhenjia Zhou described. This is especially true for agentic models like Kimi K2, where performance depends as much on the model as on the orchestration of the tools around it.

@cite:comment-evaluer-un-systeme-llm

SECTION 6

From Kimi K2 to Kimi K3: Moonshot AI's Family Evolves Through 2026

Since this article was first published, Moonshot AI hasn't slowed its release cadence, showing just how quickly an LLM's window of relevance closes in this industry.

### K2.5 and K2.6: multimodality and cost gains

In January 2026, Moonshot AI released Kimi K2.5, a multimodal upgrade to Kimi K2 featuring a native vision encoder called MoonViT, capable of processing both images and video, even replicating a user's journey through a website from a single video demonstration. Kimi K2.6, released in April 2026, then refined coding performance to match leading competing models on benchmarks such as SWE-Bench Pro, while continuing to drive down cost per million tokens.

### Kimi K3: Moonshot AI's new flagship model

The K2 series has since given way to Kimi K3, Moonshot AI's new flagship model, with a parameter count significantly larger than its predecessors. It immediately ranked among the very top models on general-purpose leaderboards, while standing out especially on front-end development benchmarks. This trajectory confirms this article's central thesis: the diversity and velocity of Chinese open models are no longer a one-off anomaly but a durable feature of the generative AI landscape.

SECTION 7

Openness, Sovereignty, and Market Dynamics

Beyond Kimi K2's individual case, the proliferation of competitive open-weight models is changing the game for enterprises. It opens the door to self-hosting for organizations most concerned with privacy or data sovereignty, to a more diverse inference provider landscape, and to downward pressure on API prices that competition between Western and Chinese players has sustained across several release cycles.

This dynamic also has concrete effects on how development teams integrate generative AI day to day: choosing a model is no longer a one-time decision, and the ability to quickly switch between providers, or combine several models depending on the task, is itself becoming a productivity lever for coding assistants.

@cite:a-quelle-vitesse-les-assistants-de-codage

SECTION 8

Conclusion

Kimi K2 will be remembered as one of the milestones that established Chinese open models as a mainstream fixture of the generative AI landscape, alongside DeepSeek. Its mixture-of-experts architecture, agentic positioning, and cost-to-performance ratio made it a relevant choice for specific use cases, particularly parallelizing coding tasks on a tight budget, at the cost of clear trade-offs on latency and context window.

Its rapid succession by K2.5, K2.6, and then Kimi K3 is a reminder of a simple rule for any organization evaluating LLMs: today's conclusion has a shelf life, and choosing a model should always rest on continuous evaluation rather than a fixed decision.

FAQ

Frequently asked questions