Author: Climber, CryptoPulse Labs

On July 16, the Dark Side of the Moon officially launched the new generation open-source model Kimi K3. The model has 2.8 trillion parameters, a 1 million Token context window, native support for visual understanding, and employs technologies such as Kimi Delta Attention and Attention Residuals.

This is the world's first open-source model at the 3 trillion parameter level. Although Kimi K3's overall performance still lags behind the strongest closed-source models like Claude Fable 5 and GPT-5.6 Sol, it displays cutting-edge levels in multiple evaluations, and the Dark Side of the Moon claims its overall performance is stable and surpasses other tested models.

What is even more noteworthy is that Kimi K3 has also independently completed the design of a chip. A large model begins to attempt to design the hardware needed to run AI, which may be more notable than the 2.8 trillion parameters themselves.

1. Behind the 2.8 trillion parameters: The competition for large models shifts from scale to efficiency

In the past few years, the most easily understandable metric in the large model industry has been the number of parameters. From tens of billions to hundreds of billions, then to trillions, parameter scale has almost become an important symbol for measuring model capability.

But once the model scale reaches 2.8 trillion parameters, the real question is no longer "how large the model is," but rather: how to train such a massive model? How many parameters are actually needed for each task? How to enhance the model's capabilities while controlling operational costs?

Kimi K3's answer is to further expand the sparse architecture.

According to the Dark Side of the Moon, Kimi K3 adopts the Mixture of Experts architecture. The model contains 896 expert modules, but only activates 16 of them for each task.

This means that the model can have vast knowledge capacity without the need to call all parameters each time. It's like a super institution with 896 specialized departments that only mobilizes the 16 most relevant departments when faced with different issues.

The core value of this architecture lies in the separation of total scale and single computation costs.

In the future, the competition among large models may not be about who has more parameters, but rather who can effectively call more parameters at a lower cost.

An additional core innovation of Kimi K3 is Kimi Delta Attention, or KDA. Traditional Transformer architectures face significant computational load and memory pressure when processing long texts. The goal of KDA is to improve the model's efficiency in handling long sequence information.

At the same time, Kimi K3 introduces Attention Residuals, a mechanism for attention residuals. Traditional models typically transmit information gradually across layers, accumulating information in subsequent layers, but this can lead to redundancy and attenuation.

Attention Residuals attempt to allow the model to selectively access information from earlier stages across different depths.

If traditional model information flow resembles a river flowing from the starting point to the endpoint, then Attention Residuals are more like establishing an information retrieval system along the way, allowing the model to call upon information from different depths based on the task.

The Dark Side of the Moon states that compared to Kimi K2, Kimi K3 achieved an overall efficiency improvement of about 2.5 times.

This indicates that the AI industry is shifting from "the larger the scale, the stronger" to "how to convert a larger scale into higher efficiency."

The significance of Kimi K3 is not merely the introduction of a 2.8 trillion parameter model but also the further elevation of the scale limits of open-source models.

In the past, open-source models were often seen as followers of closed-source models. Now, open-source models are starting to attempt to prove that ultra-large scale models can also be open, studied, and re-developed.

2. From chatbots to digital employees: Kimi K3 targets complex work

If the 2.8 trillion parameters are the most easily spread label of Kimi K3, then its real product direction is long-term tasks.

In the past, AI assistants mostly answered questions. The user asked a question, and the model provided an answer. If asked to write code, it would return code. If asked to summarize an article, it generated a summary.

But in reality, complex work often cannot be completed in one Q&A session.

A researcher may need to read papers, organize data, build models, run experiments, analyze results, and then write reports. A programmer may need to read numerous documents, understand project structures, modify code, run tests, locate bugs, and iterate continuously.

The common characteristics of these tasks are long cycles, multiple steps, and large amounts of information, requiring continuous adjustments of next actions based on intermediate results, which is precisely the issue Kimi K3 attempts to address.

In a case demonstrated by the Dark Side of the Moon, Kimi K3 completed a computational astrophysics research task. By reading and cross-verifying over 20 papers, it conducted numerical calculations, completed evaluations of hundreds of state equations, identified inconsistencies in published formulas, and generated over 3000 lines of Python code and an interactive HTML dashboard.

The official claim is that this task took approximately two hours, whereas in traditional circumstances, it might require one to two weeks for an experienced researcher to finish.

This does not mean that AI can replace researchers; the most important aspects of research work typically involve posing questions, judging hypotheses, and interpreting results.

However, Kimi K3 demonstrates an important change: AI is evolving from assisting humans in completing certain steps to gradually autonomously completing an entire workflow. This marks the distinction between the Agent era and the traditional chatbot era.

Traditional chatbots manage "what you ask, I answer." Agents, however, handle "you tell me your goal, and I break down the tasks, call tools, execute steps, check results, and continually make corrections."

Kimi K3's 1 million Token context window is significant in this process.

For large code repositories, research reports, corporate materials, and complex project documents, the model's ability to understand more information at once means it doesn't need to frequently forget context or require users to repeatedly provide background.

At the same time, Kimi K3 natively supports visual understanding, allowing AI to form a more complete work loop.

For example, after the AI writes code, it can check the results on the webpage; after creating a PPT, it can inspect the page layout; after generating content, it can determine outcomes through visual feedback.

Past AI resembled writing code with closed eyes, while future AI can form a cycle of understanding tasks, generating results, observing outcomes, identifying issues, and modifying results.

The Dark Side of the Moon has also extended Kimi's capabilities to scenarios like Kimi Work, Kimi Code, and Kimi API, focusing on research, documentation, slides, spreadsheets, dashboards, and complex programming tasks.

In the future, truly commercially valuable AI might not be the model that answers the most questions but rather the model that completes the most tasks.

Traditional software requires users to learn complex operational processes, whereas AI Agents aim to connect search, databases, programming, data analysis, and office tools so users only need to describe the final goal.

This implies that future competition in the software industry may not be about who has more tools, but rather who has a stronger AI execution system.

3. The most notable aspect is not the model, but AI's start in chip design

The most shocking part of Kimi K3 may be its independent completion of chip design.

According to information disclosed by the Dark Side of the Moon, Kimi K3, during a 48-hour autonomous run, used open-source EDA tools and the Nangate 45nm process library to complete a chip design, optimization, and verification aimed at its own architecture for small models.

This does not mean Kimi K3 can independently complete the commercial mass production of modern advanced process AI chips. The 45nm process has a huge gap compared to today's most advanced AI accelerators, and chip production involves a complex IP, process, manufacturing, packaging, and supply chain system.

However, this attempt remains significant because chip design is not merely about writing code; it requires handling multiple aspects like logical design, synthesis, layout routing, timing analysis, power optimization, and physical verification.

In the past, AI in the chip industry mainly assisted engineers in completing local tasks, such as optimizing layouts, predicting timing, and identifying design flaws.

What Kimi K3 is showcasing is another possibility—AI is no longer just using tools, but is beginning to autonomously organize tools to complete the entire engineering process.

This development path is very similar to that of AI writing code. Early AI could only generate small snippets of code, later it could write complete programs, and after that, it was able to read codebases, run tests, and fix bugs. Now, AI is starting to attempt to design the hardware needed to run AI.

This may create a new self-reinforcing loop for AI: AI helps design better chips, stronger chips train more powerful models, and more powerful models assist in designing the next generation chips.

More notably, Kimi K3 has also demonstrated the capability to independently develop GPU programming systems.

The Dark Side of the Moon disclosed that Kimi K3 developed MiniTriton, a compact compiler system similar to Triton, which includes its own intermediate representation layer, optimization processes, and PTX code generation processes.

This indicates that the boundaries of AI capability are expanding from using software to creating software tools.

In the future, the model itself might directly participate in chip optimization, compiler development, operator adaptation, and system tuning, which could be Kimi K3’s most important strategic value.

It is not just a model product, but is exploring an AI-native research and development model. From models to compilers, from algorithms to chips, from data to applications, AI gradually becomes a part of the entire infrastructure.

Of course, AI-designed chips still require strict validation, and AI-generated research results need to be reviewed by professionals; AI may also make errors when autonomously executing complex tasks.

Nevertheless, Kimi K3 has released an important signal: AI is gradually transitioning from an object of creation to a subject participating in the creation of the next generation of AI.

Conclusion

The release of Kimi K3, on the surface, is an upgrade of the model, but behind it represents a change in the logic of large model competition.

From larger parameter scales to more efficient architectures. From answering questions to completing complex tasks, and then to autonomously developing compilers and designing chips, AI is gradually participating in the creation of the next generation of AI.

2.8 trillion parameters may just be a number, but what is truly worth noting is that AI is starting to attempt to design its own future.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。