What is Kimi K3? Deep Dive into Moonshot's 2.8 Trillion Parameter AI Model

Recently, we have started to see AI models competing beyond just answering questions or helping write short snippets of code. The new goal is shifting towards building AI Agents capable of handling complex tasks, performing multi-step operations continuously, utilizing external tools, and verifying their own results while working.

One of the models currently gaining attention is Kimi K3, the new flagship model from Kimi, developed by Moonshot AI. The company stated that this model has a total of 2.8 trillion parameters, supports a context window of up to 1 million tokens, possesses native vision capabilities within the model, and is specifically designed for long-horizon coding, knowledge work, and reasoning. However, the figure of 2.8 trillion parameters does not mean that Kimi K3 will be the absolute best in every aspect, nor does it use all parameters simultaneously every time. Behind the scenes, there is a Mixture of Experts architecture, the method of selecting experts, and evaluation conditions that need to be collectively understood.

This article will take you through what Kimi K3 is, how its structure differs from typical models, its strengths in coding and agentic tasks, as well as the limitations that developers should know before testing it out.

Note: The information in this article was verified as of July 22, 2026. Moonshot announced that it will release the full model weights by July 27, 2026, and will publish an additional Technical Report around the same time.

What is Kimi K3?

Kimi K3 is the new flagship AI model from Kimi, which Moonshot stated is the company's most capable model at the time of its launch.

The main specifications revealed by the company include:

A total of 2.8 trillion parameters
Context window up to 1 million tokens
Native vision capabilities built into the model
Support for long-horizon coding tasks requiring continuous work over time
Support for end-to-end knowledge work and content generation
Designed for tasks requiring reasoning and agentic workflows

Moonshot calls Kimi K3 the first open model to enter the 3-trillion parameter tier, or a 3T-class model. However, as of the date this information was verified, the full model weights were not yet available for download. The company has scheduled the release for July 27, 2026, along with further details regarding its architecture, training, and evaluation.

Therefore, when discussing the status of Kimi K3 during its launch, two things should be separated: Moonshot has announced Kimi K3 as an open model, but the full model weight files are still in the process of being prepared for release according to the company's schedule.

2.8 Trillion Parameters

The figure of 2.8 trillion parameters might lead many to understand that Kimi K3 must use all parameters for computation every time it generates a response. But its actual structure does not work that way.

Kimi K3 utilizes a Mixture of Experts (MoE) architecture. Internally, a portion of the model is divided into 896 experts, and it selectively activates approximately 16 experts during each processing phase through a system called Stable LatentMoE. Imagine having a large corporation with hundreds of expert teams. When a task comes in, the system doesn't need to call every employee to work simultaneously; it will only select the teams best suited for that specific problem to take responsibility. This principle allows the model to possess a massive overall scale of knowledge and capability, while the processing of each token utilizes only a fraction of the model.

However, the total number of parameters does not dictate the entire performance. A model with more parameters might not be faster, more cost-effective, or provide better answers across all types of tasks. What must be evaluated together includes the architecture, training methodology, data quality, tool utilization, and the nature of the tasks being tested.

Key Architecture of Kimi K3

Moonshot stated that Kimi K3 was developed upon several key architectural components to enable the model to better handle long contexts and large-scale structures.

1. Kimi Delta Attention

Kimi Delta Attention, or KDA, is an attention mechanism used as one of the foundations of Kimi K3. Official API documentation describes KDA as a hybrid linear attention mechanism. The role of attention is to help the model consider relationships between different pieces of data, such as the relationships between functions across multiple source code files, content within documents, or text that occurred in a previous conversation. As the context length increases, the computational load of attention also increases accordingly. Therefore, KDA was introduced as a foundation to expand the capacity for handling longer sequences.

2. Attention Residuals

Attention Residuals, or AttnRes, were designed to improve information flow through the model's depth. Moonshot explains that AttnRes can selectively draw representations from various depths instead of merely accumulating data from previous layers in a uniform manner throughout the model. This concept allows a large-scale model to more effectively select appropriate data from different parts of its structure.

3. Stable LatentMoE

Stable LatentMoE is the component that helps manage Kimi K3's 896 experts. The model will activate approximately 16 experts during each processing phase. Moonshot also noted that it uses Quantile Balancing to allocate tasks to experts based on router scores, reducing the issue of certain experts receiving too much or too little work. By combining architectural changes with adjustments to the training process and data, Moonshot estimates that Kimi K3 has an overall scaling efficiency roughly 2.5 times higher than Kimi K2. However, this figure is an evaluation result disclosed by the model developers themselves.

What Can a 1 Million Token Context Window Help With?

Kimi K3 supports a context window of up to 1 million tokens, meaning the model can take in a massive amount of data within a single session.

For developers, this capability might be ideal for tasks that require considering large volumes of information simultaneously, such as:

Reading source code from multiple files
Analyzing large repositories
Reviewing code alongside documentation
Analyzing massive amounts of logs
Comparing multiple versions of code or documents
Executing continuous multi-step tasks using data from previous steps
Using internal documents as context for a coding agent

However, a large context window does not mean we should dump all data into it without organization. If the majority of the information is irrelevant to the prompt, the model might consume more tokens and time without proportionally improving the outcome. Therefore, designing a structured context, prioritizing critical data, and filtering out irrelevant information remains something developers must continue to focus on.

Key Strengths of Kimi K3

One of the capabilities Moonshot emphasizes the most is Long-horizon Coding, or software development tasks that don't just end with generating a single function's code but require continuous multi-step work. Moonshot stated that Kimi K3 can operate within long engineering sessions, navigate large repositories, and coordinate seamlessly with terminal tools while requiring a lower level of human supervision compared to typical coding tasks.

An example workflow might consist of:

Reading requirements and exploring project structures
Searching for related files or functions
Planning the modifications
Editing source code across multiple files
Running build or test commands
Reading any resulting errors
Going back to fix the code again
Verifying the results before finalizing the task

This highlights the difference between an AI that assists with code completion in chunks versus an Agent attempting to take responsibility for a much longer development process.

Vision in the Loop

Kimi K3 does not only analyze text and source code. Moonshot indicated that it can also utilize screenshots and visual feedback alongside software development. The company refers to this approach as "Vision in the Loop," which means allowing the model to alternate between writing code, viewing the output, and using the observed images to further refine the work.

For example, if Kimi K3 is building a webpage, the agent might work in the following sequence:

Creating the HTML, CSS, and JavaScript structure
Opening the generated webpage
Checking the screenshot of the output
Identifying any misplaced layouts or elements
Going back to edit the source code
Opening and reviewing the results again

Moonshot stated that capabilities of this nature can be applied to frontend development, game development, CAD, and tasks that require a combination of software engineering and visual reasoning. However, performance in real-world projects still depends on the tools the agent can access, the environment used to run the code, and the clarity of the requirements.

Coding Task Examples Demonstrated by Moonshot

On Kimi K3's launch page, Moonshot presented several case studies to demonstrate how the model can handle long and complex engineering tasks.

1. GPU Kernel Optimization

Moonshot tested the model by having it optimize a GPU kernel within the same sandbox, giving it a maximum of 24 hours to analyze, rewrite, and benchmark various types of workloads. From the results disclosed by the company, Kimi K3 performed competitively against Claude Fable 5 under the specified conditions and scored higher than some other compared models. But this result should not be used to conclude that Kimi K3 writes code better than other models in every situation. The testing details noted that each model might have used a different Agent Harness, and Claude Fable 5's results came from third-party evaluations which might include fallback behaviors.

2. Building a GPU Compiler

Moonshot stated that Kimi K3 built a system named MiniTriton, a lightweight compiler with an approach similar to Triton, comprising an Intermediate Representation, Optimization Passes, and a PTX code generation process. The company tested MiniTriton against supported benchmarks, including end-to-end training of nanoGPT, and reported that the resulting loss curve trended closely to the reference system. This data comes from an experiment disclosed by Moonshot itself, so it should be viewed as a case study showcasing the agent's potential, not as definitive proof that Kimi K3 can build production-level compilers without engineer oversight.

3. Chip Design

Another example involved having Kimi K3 design a chip for processing a nano-model built on its own architecture. Moonshot noted that the agent worked continuously for about 48 hours, using open-source Electronic Design Automation (EDA) tools and verifying results through simulation. The company presented this project as a proof of concept for its long-horizon agentic work capabilities. Therefore, the crucial point of this example isn't a declaration that AI can now replace chip design engineers, but rather a demonstration that the model can chain together multiple steps of a workflow and operate continuously over extended periods.

Kimi K3 in Research and Knowledge Work

Kimi K3 is not positioned solely for coding. It is also aimed at end-to-end knowledge work—tasks that require gathering data, analyzing it, generating content, and presenting the results within the same workflow.

Moonshot presented examples of using Kimi K3 for various types of tasks, such as:

Analyzing industry data and building interactive websites
Generating reports featuring graphs and visualizations
Analyzing scientific research papers
Creating editable presentations and infographics
Processing massive amounts of data using multiple sub-agents
Creating motion graphics and editing videos

The company stated that Kimi K3 is a natively multimodal model, capable of understanding text, images, and videos within the same model, and applying that understanding to motion design, animation, and video editing tasks.

However, the volume of documents, the working duration, and the quality of the outcomes in all examples are data from Moonshot's case studies. Thus, they should be used to understand the scope of capabilities the company wishes to showcase, and should not be taken as a guarantee of achieving the same level of results across all projects.

Is Kimi K3 Better Than Claude and GPT?

The answer is that such a conclusion cannot be drawn just yet. Moonshot candidly stated that while Kimi K3 performs at a frontier level in the company's evaluation suites, its overall performance still trails behind some of the most powerful proprietary models, namely Claude Fable 5 and GPT-5.6 Sol. Kimi K3 might score well in specific benchmarks especially in coding, agentic workflows, and tasks requiring continuous processing but no single model necessarily has to win in every category of work.

Reading benchmarks also requires considering additional details, such as:

Which Agent Harness was used during testing
What reasoning level was enabled
What hardware was used
Whether fallback systems were in place
Whether scores were sourced from the developers themselves or external entities
Whether tests were run once or averaged over multiple runs
Whether the evaluation conditions were identical for all models

In the footnotes of the launch page, Moonshot specified that Kimi K3's results utilized a "max" reasoning effort, and each benchmark suite might evaluate the model differently via Kimi Code, Claude Code, or Codex.

Therefore, for development teams, a more reliable method than choosing based on graph rankings is to test the model against real-world repositories, prompts, tools, and system requirements.

Where Can You Access Kimi K3?

Moonshot provides access to Kimi K3 through several main product channels:

Kimi.com
Kimi Work
Kimi Code
Kimi API

For Kimi Work, the company requires using the Desktop app version 3.1.0 or higher, supporting Windows and Apple Silicon Macs. As for Kimi Code, Kimi K3 can be selected via the /model command within the terminal.

Developers wishing to integrate the model into their own systems can call it via the Kimi API using the model name kimi-k3. Official documentation includes examples of making calls via the OpenAI SDK by configuring Moonshot's Base URL and API Key.

Kimi K3 API Pricing

Prices announced by Moonshot for the Kimi K3 API (updated July 22, 2026):

Cache-hit Input: $0.30 USD per 1 million tokens
Cache-miss Input: $3.00 USD per 1 million tokens
Output: $15.00 USD per 1 million tokens

Moonshot also stated that the official API achieves a cache hit rate of over 90% on the company's coding workloads. However, the actual cache hit rate in any given system may vary depending on the specific context and prompt structures used.

Who is Kimi K3 Suitable For?

Based on the capabilities revealed by Moonshot, Kimi K3 is likely suited for complex tasks requiring continuous, multi-step execution, such as:

Development teams needing to analyze massive codebases
Teams building coding agents
Developers who want AI to collaborate seamlessly with the terminal
Frontend or game development tasks using screenshots as feedback
Research work requiring simultaneous document reading and code generation
Data analysis and visualization generation tasks
Workflows combining text, images, and video
Organizations with the infrastructure to experiment with large-scale models

However, for general Q&A, short text summarization, or systems requiring low latency, smaller models might be more appropriate and easier to control costs.

Thus, choosing a model shouldn't rely solely on parameter counts or benchmark scores. Instead, you should evaluate whether the model's quality, speed, cost, and behavior fit your specific workflow.

Frequently Asked Questions About Kimi K3

How many parameters does Kimi K3 have?

Kimi K3 has a total of 2.8 trillion parameters. It uses a Mixture of Experts architecture featuring 896 experts, activating approximately 16 of them during each processing phase.What is the maximum context window supported by Kimi K3? Kimi K3 supports a context window of up to 1 million tokens, making it ideal for processing codebases, documents, or massive amounts of data within a single session.

Is Kimi K3 an open-source model yet?

Moonshot refers to Kimi K3 as an open model and a 3-trillion parameter open-source model. However, as of July 22, 2026, the full model weights are not yet available for download. The company announced they will be released by July 27, 2026.

Can Kimi K3 be used for writing code?

Yes. Kimi K3 is designed for long-horizon coding. According to Moonshot's revealed capabilities, it can work with large repositories, coordinate terminal tool usage, and utilize screenshots as feedback in frontend, game development, and CAD tasks.

Is Kimi K3 better than Claude or GPT?

It cannot be concluded that it is better in every aspect just yet. Moonshot itself stated that Kimi K3's overall performance still trails behind Claude Fable 5 and GPT-5.6 Sol, even though it performs well in certain benchmarks, particularly in coding and agentic workflows.

Can I run Kimi K3 on my personal computer?

The full model weights are extremely large, and Moonshot recommends deploying it via a Supernode with at least 64 accelerators. Therefore, it is not a model suited for typical personal computers. Most users will likely access it via services or APIs.

Does Kimi K3 support the Thai language?

The launch page and official API documentation used as sources for this article did not publish specific Thai language benchmarks. Thus, it shouldn't be prematurely concluded what its proficiency level in Thai is. If you intend to use it practically, it should be tested directly against the specific datasets and language patterns of your system.

Conclusion

Kimi K3 represents another major milestone in large AI model development, boasting 2.8 trillion parameters, a 1-million token context window, and a Mixture of Experts architecture that selectively activates approximately 16 out of 896 experts during each processing phase. But what is more fascinating than the model's size is the direction Moonshot is pushing Kimi K3 towardserving as an AI Agent for long-horizon tasks. This includes navigating repositories, editing multiple files, using terminals, reviewing screenshots, analyzing research, and generating end-to-end output. Simultaneously, Moonshot has disclosed critical limitations, including sensitivity to thinking history, potentially excessive proactive behavior, and UX gaps when compared to top-tier proprietary models.

Therefore, Kimi K3 shouldn't be viewed as the best model simply because it has the most parameters. Instead, it should be seen as an example of a new direction where open models are shifting from mere chatbots or code assistants to systems capable of taking on software engineering workflows and complex knowledge work for developers. The most critical question isn't just which model is larger or scores higher on benchmarks; it's which model can execute your tasks accurately, remain controllable, integrate seamlessly with your existing tools, and offer the most cost-effective solution for your actual system.