The global artificial intelligence race has gained another major contender. Kimi K3, developed by Chinese AI company Moonshot AI, is a new large-scale model designed for reasoning, coding, long-form knowledge work, multimodal tasks, and autonomous AI agents.
Released in July 2026, Kimi K3 immediately attracted attention for one extraordinary number: 2.8 trillion total parameters. Rather than simply making a conventional language model larger, Moonshot AI combined enormous scale with a Mixture-of-Experts architecture, a million-token context window, native visual understanding, and new attention mechanisms intended to make that scale more efficient.
Kimi K3 is particularly significant because its model weights have been released publicly, giving developers much more freedom to study, deploy, modify, and build upon the system than with most proprietary frontier models.
What Is Kimi K3?
Kimi K3 is a frontier-scale artificial intelligence model created by Moonshot AI, the Beijing-based company behind the Kimi family of AI assistants.
The model is designed for much more than ordinary chatbot conversations.
Moonshot positions K3 for tasks including:
- Advanced reasoning
- Long-horizon software development
- Research and knowledge work
- Document analysis
- Visual understanding
- AI agent workflows
- Tool use
- Complex multi-step projects
The company describes K3 as its most capable model so far.
Its goal is increasingly to complete projects rather than simply answer individual questions.
Why 2.8 Trillion Parameters Does Not Mean All 2.8 Trillion Are Used at Once
Kimi K3 contains approximately 2.8 trillion parameters, placing it in the three-trillion-parameter class.
That number sounds enormous because it is.
However, K3 uses a Mixture-of-Experts, or MoE, architecture.
Instead of activating the entire network for every piece of input, the model contains many specialized neural-network components known as experts. A routing system chooses a small subset of them for each token.
Kimi K3 uses a framework with 896 experts while activating 16 at a time. This allows the model to have enormous total capacity without requiring every parameter to participate in every computation.
This architecture helps explain how modern AI developers can continue increasing model capacity without computational costs rising exactly in proportion to total parameter count.
A One-Million-Token Context Window
Another standout feature is Kimi K3’s one-million-token context window.
A context window determines how much information an AI model can consider during a task.
Large context windows can be especially useful for working with extensive source material, including large code repositories, long research documents, collections of reports, and complicated multi-stage projects.
A million tokens potentially allows K3 to process information far beyond the length of an ordinary book.
But long context does not automatically guarantee perfect memory.
Models still need to identify the most important information, connect distant details correctly, and avoid losing accuracy as the amount of material grows.
Kimi Delta Attention
Scale is only part of K3’s design.
Moonshot AI introduced an architecture based on Kimi Delta Attention, commonly shortened to KDA.
Attention mechanisms are fundamental to modern transformer-based AI because they help models determine which pieces of information matter when generating the next output.
Traditional attention becomes increasingly expensive as sequences become very long.
KDA is designed to improve the efficiency of processing long sequences while preserving strong model capabilities.
K3 also uses Attention Residuals, another architectural mechanism designed to improve how information travels through the model’s many layers. Moonshot says these architectural changes and other improvements substantially increased scaling efficiency compared with earlier Kimi generations.
Native Vision Makes K3 More Than a Text Model
Kimi K3 has native visual capabilities.
That means images and visual documents are part of the model’s core multimodal design rather than being treated purely as an external add-on.
A multimodal model can potentially analyze combinations of text and imagery such as:
- Diagrams
- Screenshots
- Charts
- Documents
- User interfaces
- Technical drawings
- Photographs
This becomes especially valuable for agentic AI.
A coding agent, for example, may need to inspect a webpage visually, recognize that a button is misplaced, modify the code, run the application again, and inspect the result.
Kimi K3 Is Designed for Long-Horizon Coding
Coding is one of the areas where Moonshot has placed particular emphasis.
Traditional coding assistants usually respond to relatively contained requests: write a function, explain an error, or generate several lines of code.
Long-horizon coding is different.
An AI system may need to inspect thousands of files, understand an unfamiliar project, modify multiple components, run tools, debug failures, and continue working through dozens of steps.
Kimi K3 was specifically designed to sustain these longer workflows.
Moonshot has highlighted applications ranging from ordinary software engineering to compiler work, GPU optimization, interactive development, and other highly technical tasks.
Independent Tests Show Why K3 Is Getting Attention
Independent evaluations have also helped attract attention to the model.
Artificial Analysis placed Kimi K3 among high-performing frontier systems on its intelligence evaluations and found particularly strong results on complex agentic knowledge-work tasks.
Its testing also revealed an important trade-off: K3 could use many turns and large amounts of generated text while completing difficult jobs, making some tasks relatively slow and computationally expensive.
That distinction matters.
An AI model can be extremely capable without necessarily being the fastest or cheapest model for every task.
For simple questions, such a huge model may be unnecessary. Its advantages become more relevant when the job requires extended reasoning, tools, coding, large amounts of context, or many interconnected steps.
Open Weights Could Be Just as Important as Performance
Moonshot released the full Kimi K3 model weights on July 27, 2026, together with technical materials and infrastructure technology associated with the project.
This gives researchers and companies the possibility of running or adapting the model outside Moonshot’s consumer application, although the hardware requirements of a 2.8-trillion-parameter system remain enormous.
Its license permits broad activities including using, modifying, distributing, fine-tuning, and creating derivative works, subject to the license conditions.
For the AI ecosystem, this matters because open-weight frontier models allow independent researchers to examine architectures and build specialized systems without depending entirely on a closed API.
Expert Perspective: Chinese Open Models Are Closing the Gap
Independent analysis of the global AI market increasingly points to the rapid improvement of Chinese open-weight models.
Artificial Analysis found Kimi K3 competitive with several leading proprietary systems across intelligence and agentic-work benchmarks, while Moonshot itself acknowledges that the model does not dominate every comparison.
The broader trend is significant.
Models from Chinese laboratories are no longer interesting primarily because they are cheaper alternatives. They increasingly compete in coding, reasoning, multimodal work, and AI-agent capabilities.
Kimi K3 demonstrates how quickly advanced open-weight AI is moving toward the performance range traditionally associated with closed frontier systems.
Why Kimi K3 Could Matter Beyond China
Kimi K3 represents several important trends happening simultaneously.
AI models are becoming more multimodal.
Context windows are becoming dramatically larger.
Mixture-of-Experts architectures are allowing developers to increase model capacity more efficiently.
AI systems are shifting from short conversations toward longer autonomous workflows.
And powerful model weights are increasingly being released for developers to build upon.
Kimi K3 combines all of these trends in one system.
Its enormous scale also exposes a limitation: running frontier open models locally is very different from downloading a normal application. Even when weights are publicly available, the computing infrastructure required for efficient deployment can be substantial.
The next phase of the competition may therefore depend not only on who can build the smartest model, but also on who can make frontier intelligence efficient enough to deploy widely.
Interesting Facts
- Kimi K3 contains approximately 2.8 trillion total parameters.
- It uses a Mixture-of-Experts architecture rather than activating the entire model for every token.
- Its architecture includes 896 experts, with 16 selected during processing.
- Kimi K3 supports a context window of up to one million tokens.
- The model includes native visual understanding.
- Moonshot AI designed K3 for long-horizon coding and knowledge-work tasks.
- Kimi K3 uses architectural techniques called Kimi Delta Attention and Attention Residuals.
- The complete model weights were released in July 2026.
- A model can have trillions of total parameters while using only a fraction of them during each computational step.
- Independent evaluations have found K3 particularly competitive on complex multi-step and agentic tasks.
- Its enormous size means that open weights do not automatically make the full model easy to run on ordinary consumer hardware.
Glossary
- Kimi K3 — A 2.8-trillion-parameter multimodal AI model developed by Moonshot AI.
- Moonshot AI — A Chinese artificial intelligence company developing the Kimi family of models and AI products.
- Parameter — A learned numerical value inside a neural network that helps determine how the model processes information.
- Large Language Model (LLM) — A neural network trained on large amounts of data to understand and generate language and perform related tasks.
- Mixture of Experts (MoE) — An architecture containing multiple specialized neural-network components while activating only selected experts for each input.
- Token — A unit of information processed by a language model, often representing part of a word, a whole word, or another small piece of data.
- Context Window — The amount of information a model can consider during a single ongoing task.
- Kimi Delta Attention (KDA) — An attention architecture used by Kimi K3 to process information efficiently, particularly across long sequences.
- Attention Residuals — An architectural technique intended to improve information flow through the layers of a neural network.
- Multimodal AI — Artificial intelligence capable of processing more than one type of information, such as text and images.
- AI Agent — An AI system designed to perform multi-step tasks using tools, software, or external environments.
- Long-Horizon Coding — Software-engineering work requiring an AI system to continue planning, editing, testing, and debugging over many steps.
- Open Weights — Publicly available trained model parameters that can be downloaded and used under the accompanying license.
- Inference — The process of running a trained AI model to generate outputs from new inputs.
