SYSTEM NOTICE

Auto translation by AI. Be sure, accuracy, nuances and authorial intent may not be fully reflected.
見出し画像

[Kini K3] "The most detailed explanation anywhere!!" A 2.8 trillion parameter open-weight AI has arrived. How will individual development change in the era of Claude Code/Codex? AI side hustle/side job/ClaudeCode/CodeX/ChatGPT/Gemini/monetization/generative AI/individual development/apps

AI かわき｜世界一優しいAI画像生成の解説


Introduction

Hello, this is AI Kawaki.

The other day, some incredible news came out that shook the industry!!

Moonshot AI has officially released the Kimi K3 model itself, meaning the trained weights and the technical report.

This is not just another new model announcement.

It is common to hear that a model is now available via API.

But this time, Kimi K3 has been released on Hugging Face and GitHub in a way that allows third parties to verify it.

An open-weight AI has reached a level close to Claude and GPT.

This is the biggest point this time.

Moreover, what Kimi K3 is aiming for is not just ordinary chat responses.

An AI that reads large codebases, operates terminals, looks at screens, fixes failures, and carries out long-term tasks.

In other words, this is quite relevant for those using Claude Code, Codex, Cursor, Cline, and Hermes-based agents.

In this article, I will organize what has changed with Kimi K3 and how it can be used for individual development and AI side hustles from a practical perspective.

Table of Contents

  • What happened with Kimi K3

  • What is so amazing about it

  • Is it stronger than Claude or GPT

  • Can it be run on a local PC

  • The most realistic way to use it

  • How to pivot this for side hustles and monetization

  • What to do if you want to try it first

  • Points to note

  • Conclusion

What happened with Kimi K3

First, here are the facts.

Kimi K3 is a 2.8 trillion parameter class open-weight model released by Moonshot AI.

The official GitHub describes Kimi K3 as an open-weight, native multimodal, agent-oriented model and the highest-performing model in the Kimi series.

Reference URL:

Official blog:

Hugging Face:

What is important here is that it is not just an API.

By releasing the model weights themselves, a foundation has been created for external researchers and developers to verify, modify, and operate it independently.

This is quite significant in the AI industry.

This is because while closed models are convenient, the freedom to verify their contents or deploy them in full-scale in our own environments is limited.

On the other hand, although Kimi K3 is too massive for just anyone to run easily, it is significant as a model that approaches the top class of open weights.

This is quietly a big deal.

From the perspective of AI side hustles and personal development, rather than talking about running it on a home PC immediately, it is more realistic to see it as an increase in powerful alternatives to Claude and GPT as brains for agent development.

What is so amazing about it

The specs of Kimi K3 are quite flashy.

The main specifications that can be confirmed on the official GitHub and Hugging Face are as follows.

Total parameter count: 2.8 trillion
Active parameters actually used: 104 billion
Expert models: 896, with 16 experts selected per token
Context length: approx. 1 million tokens, Quantization: MXFP4 weights / MXFP8 activations

Reference URL:

When you hear 2.8 trillion parameters, it sounds like everything is running at full capacity every time, but it is actually an MoE, or Mixture of Experts model.

Roughly speaking, it is a mechanism that prepares a huge team of experts and calls only the necessary experts for each task.

Instead of running everything every time, it selects and runs the necessary areas of expertise.

This design reduces the computational load during inference while increasing the model scale.

This is the interesting part of Kimi K3.

Rather than just a high-performance chat AI, it is closer to an agent foundation that calls a huge group of experts according to the task.

If you are repurposing it for individual development, rather than for ordinary chat use, it is suitable for heavier work such as:

Reading large-scale code
Understanding long specifications
A series of workflows from research to implementation across multiple tools

It is suited for these types of heavy tasks.

Is it stronger than Claude or GPT?

It is better to look at this calmly.

On the official Kimi blog, it is written that Kimi K3 has produced quite strong results in some agent-based and automation-based benchmarks.

For example, in the table posted officially, top-tier numbers are lined up for items close to agent work and tool operation, such as BrowseComp, MCPMark-Verified, AutomationBench, SpreadsheetBench 2, and Terminal-Bench.

Reference URL:

However, you cannot jump to the conclusion that it has completely surpassed Claude or GPT.

The official blog also states that in terms of overall performance, there are parts that have not yet reached the top-tier proprietary models.

Furthermore, depending on the benchmark, the execution environments differ, such as Kimi Code, Claude Code, and Codex.

In other words, it is not just about the capability of the model itself, but a comparison that includes peripheral tools and execution harnesses in some parts.

The correct assessment this time is not that GPT or Claude are finished, but that open-weight models have come very close to the top-tier closed models.

This perspective is the least likely to be off the mark.

Personally, this is the most rewarding part.

Rather than viewing it as a battle for the strongest model, it is more useful in practice to see it as the era of using Claude, Codex, and Kimi for different purposes. has arrived.

Can it run on a local PC?

This is quite important.

To conclude, it is not realistic to run the Kimi K3 model itself locally on a typical personal PC.

For example, with a configuration like an RTX 5060, the Kimi K3 model itself is difficult to run.

Kimi K3 uses 4-bit quantization called MXFP4, but even with a simple calculation,

2.8 trillion parameters × 4 bits ≒ 1.4 TB

This is the theoretical value for the weights alone.

In reality, memory for inference, KV cache, image processing, and intermediate data during execution are also required.

The official blog also recommends a supercomputer-class configuration with 64 or more accelerators for full-scale deployment of Kimi K3.

Reference URL:

So, it is better not to misunderstand this.

Weights being publicly released does not mean they can be used comfortably for free on a home PC.

If you misunderstand this, you will be quite disappointed.

However, that does not mean the open weights have no value.

It is significant that research institutions, companies, developers with GPU clusters, and teams with inference infrastructure can verify and optimize them in their own environments.

For individuals, it is more realistic to use APIs or hosted environments than to run it locally.

The most realistic way to use it

From the perspective of a personal developer or someone doing a side hustle, this is the most realistic way to use Kimi K3.

Use Kimi K3 via API as the brain for Claude Code or custom-built agents.

Kimi officially provides an Anthropic-compatible API and has also published configuration examples for using kimi-k3[1m] from Claude Code.

Reference URL:

In other words, you can operate by using a development experience like Claude Code while swapping only the underlying model to Kimi K3.

This is quite realistic.

The official documentation introduces the following settings for Claude Code.

Set ANTHROPIC_BASE_URL to Kimi's Anthropic-compatible endpoint
Specify kimi-k3[1m] for ANTHROPIC_MODEL, set CLAUDE_CODE_AUTO_COMPACT_WINDOW to 1048576, and use it with the assumption that reasoning effort is set to max

What is important here is the possibility of bringing a 1 million token context into a Claude Code-style development flow.

For larger web apps, Unity projects, SaaS spanning multiple files, or contract projects with long specifications, having the context cut off is a significant source of stress.

Kimi K3 is leaning in a direction that significantly reduces that stress.

How should we view the API pricing?

You can check the API pricing for Kimi K3 on the official blog and API documentation.

As of July 28, 2026, the pricing for Kimi K3 indicated by the official source is as follows.

Cached input: $0.30 / 1 million tokens
Standard input: $3.00 / 1 million tokens Output: $15.00 / 1 million tokens

Reference URL:

Pricing page:

What you should look at in this pricing structure, rather than the high cost of output, is the fact that cached input is quite cheap.

For example, loading a large codebase from scratch every time becomes expensive.

However, if you are working on the same project repeatedly, a design that utilizes caching is quite effective.

It is well-suited for tasks where you leave long-form background information in place and perform repeated operations.

This is the highlight of Kimi K3 in terms of pricing.

Conversely, for use cases that require large amounts of output every time, such as generating hundreds of thousands of tokens of long text, the output cost will have an impact.

If you are using it for a side hustle, it is easier to use by focusing on design, research, code comprehension, and creating correction plans.

Points to Note

Kimi K3 is quite interesting, but there are some points to note.

First, even though it is an open-weight model, it is not a model that can be run as-is on a home PC. Next, since benchmarks are primarily measured by the official team, there is no guarantee that you will get the same results in actual projects.

Furthermore, in use cases like Claude Code integration, there are features that Kimi's Anthropic-compatible API does not support.

The official documentation also explains that since Claude Code's Tool Search is not yet supported by the Kimi endpoint, you need to set ENABLE_TOOL_SEARCH to false.

Reference URL:

This is important.

Just because it runs on the Claude Code interface does not mean it will behave exactly the same as Claude.

Kimi K3 is a very strong option.

However, before handing over an existing project entirely, you should always test it with a copy for verification.

AI agents can make rapid progress when they succeed. However, when they make a mistake, they can break things just as quickly.

That is precisely why Git management, diff checking, breaking tasks into small pieces, and having a verification environment are essential.

Summary

With this Kimi K3 release, this is the most significant point.

Open-weight models have reached a point where they are close to the top-tier closed models in long-duration coding, research, and tool operation.

This is also highly relevant to AI side hustles.

This is because in future personal development and contract work, it won't be enough to just have AI write text;

researching → designing → writing code
→ verifying operation → fixing → summarizing into articles or sales materials

people who can handle this entire workflow will be the ones who excel.

Kimi K3 is a candidate for the brain that can support that workflow.

It is not a model for free operation on a local PC.

However, it is quite interesting if you incorporate it into Claude Code or your own custom agents via API.

Personally, these are the three use cases you should try first:

1. Modifying larger existing codebases
2. Research tasks involving reading long documents
3. Rapid prototyping of small-scale SaaS or web tools

These are areas where Kimi K3's long context and agent performance can be easily leveraged.

Claude, Codex, Kimi.

It's not about which one is the strongest, but rather,I believe those who can choose which brain to use for which task will become the strongest.

Official information referenced

I also post on X every day! Please be sure to follow me!

I've started a Substack!

Get the latest AI news, useful AI tools, and helpful tips for AI side hustles delivered directly to your email for free.
You can read them in a slightly more organized format than on X, so please sign up for this as well.

Membership Information

Thank you for your continued support.
I currently run a membership program where I deliver weekly member-only articles on monetization, the reality of AI side hustles, and exclusive columns.
If you are interested, please join us.

Click here for related articles ↓

Next time, I will continue to share ideas on AI image generation and prompt design.
See you in the next article.

いいなと思ったら応援しよう！

AI かわき｜世界一優しいAI画像生成の解説 いただいたチップは、より質の高いコンテンツを制作するための活動資金として、大切に活用させていただきます。