Jensen Huang, Kimi K3, and an Upcoming Open Source War

Author: Digital Life Kazik

If you were to mention the hottest and most discussed topic in the AI industry recently, it would undoubtedly be a story woven together by four seemingly unrelated words: Jensen Huang, Anthropic, Kimi, and open source. The story goes like this. A few days ago, on July 24, Jensen Huang posted his first message on X.

The most popular man in the AI circle registered for X just to send this open letter. The letter is titled "Open Weights and American AI Leadership."

The gist is that AI will transform every industry, drive every company, and will be built by every country. Open models can enhance security, accelerate innovation, and bring about technological sovereignty. The world needs both cutting-edge closed-source models and cutting-edge open models. We cannot sanction open-source models. The reason Jensen Huang is so concerned is that there is a backstory. Let's rewind to 11 days ago when a landmark event occurred. In the early hours of July 17, Kimi K3 was officially released.

With a total of 28 trillion parameters, 104 billion active parameters, native visual capabilities, and a context of 1 million tokens.

The key point is that Kimi promised at that time that the model weights would be officially open-sourced on July 27 (which was last night).

This is the world's first open-source 3T-level model. A Chinese team, despite not having ample computing power, created something that approaches the capabilities of the world's strongest closed-source models and is willing to share the weights with the world.

This immediately changed the atmosphere in the U.S. On July 22, Michael Kratsios, head of the White House Office of Science and Technology Policy, publicly accused Kimi K3 of having developed using large-scale distillation from Claude Fable 5.

Interestingly, the comments section was lively. People can see the public's outcry.

Then, the U.S. Treasury Secretary, unfortunately, added fuel to the fire by mentioning trade blacklists and sanctions.

The comments section was completely overwhelmed.

This matter escalated, and what was originally a competition about model capabilities and open-source routes was quickly dragged into issues of intellectual property, national security, export controls, and U.S.-China tech competition. At this point, Jensen Huang stepped up.

He gave an interview and publicly opposed banning Chinese open-source models.

Two days later, he posted his first X, using an open letter to tell Washington that open models are crucial for American innovation, competition, security, and technological sovereignty; you cannot ban or sanction them, we must encourage open source.

To be fair, Jensen Huang is genuinely concerned about selling cards. Then, at the bottom, a long logo wall was displayed, and the entire tech industry began to line up to sign the open letter. When the open letter was first released, 25 companies and organizations had signed it.

NVIDIA, Microsoft, Meta, IBM, Dell, Palantir, Mistral, Hugging Face, and many more.

By the weekend, the list began to grow rapidly.

OpenAI joined, Google joined, SpaceX joined, AMD, Cisco, Cloudflare, GitHub, Linux Foundation, Mozilla, Vercel, Y Combinator, OpenClaw, and so on, all joined in.

By the time I wrote this article, there were already 77 signatories.

Numerous big names were also wildly supporting it.

It has been a long time since I have seen such a gathering, rising up together, united against a common enemy. However, careful observers will notice that the logo wall above almost gathered most of the leading tech companies in the U.S., and even NVIDIA and AMD were in the same frame, but you will find that a familiar name is inexplicably missing. Anthropic. I think this world-famous painting could be renamed "Where is Anthropic?". Some netizens even humorously made a video about it, and I was laughing so hard that I had to show it to everyone.

While Anthropic played dead, American tech giants united to rise up on Twitter. The trigger for this incident was officially announced last night, on July 27 at 23:13, that Kimi K3 was open-sourced.

Moreover, the level of open source exceeded my expectations. I originally thought it was just the model weights that would be open-sourced.

But it turned out that.

Model weights, open-sourced.

A detailed technical report was also released.

The key infra used to train K3 was also open-sourced.

Everything is open...

Let me list a few things.

Kimi K3 model weights, a 28 trillion parameter MoE model, with a total of 896 experts, activating 16 of them for each token.

Then there is the complete technical report of K3, totaling 47 pages.

This report disclosed the architectural details of KDA and Attention Residuals, explaining how they mixed KDA and Gated MLA in a 3 to 1 ratio, and how to allow information to flow across dozens of layers.

There is also Stable LatentMoE.

896 experts activate only 16, with an extremely high sparsity. Kimi used SiTU GLU and Quantile Balancing to maintain this extreme sparse training in a stable state.

There is also MoonViT V2, reinforcement learning infrastructure for a million-token context, large-scale task synthesis, post-training methods, and complete results of nearly 20 internal evaluation sets were also released.

This technical report is very detailed and solid, highly recommended for reading.

In addition to the technical report, there are also open-sourced key infra technologies.

MoonEP, a communication library specifically designed for ultra-large-scale MoE expert parallelism.

During MoE training, it often happens that some experts are overwhelmed while others are idle, causing all GPUs to wait for the busiest card.

MoonEP dynamically replicates a small number of popular experts based on the current routing results, allowing each card to receive the same number of tokens while maintaining a static memory shape, avoiding fragmentation during training and ultimately leading to OOM.

This is already something very fundamental and valuable in the training of large models.

FlashKDA, a high-performance operator of Kimi Delta Attention. It is implemented using CUTLASS and, on NVIDIA H20, the prefill speed has improved by 1.85 to 2.31 times compared to the flash linear attention baseline, and it can be directly integrated into the existing flash linear attention backend.

Next, we have AgentENV.

This is a distributed Agent sandbox system developed by Kimi and KVCache.ai, serving as the foundational layer for K3's large-scale Agent reinforcement learning environment.

It can simultaneously launch massive Firecracker micro virtual machines across multiple machines, with environment recovery times under 50 milliseconds and pauses under 100 milliseconds, and it natively supports snapshots, recovery, and forking.

Once an Agent reaches a critical node, it can directly fork into dozens of non-interfering trajectories to continue parallel exploration.

Although I am not very familiar with underlying technologies and only possess some superficial knowledge, I can already sense the excitement from these technologies.

This is truly digging deep.

MoonEP addresses the communication and load issues of ultra-large MoE models.

FlashKDA solves the operator efficiency of new attention architectures.

AgentENV provides the massive real environments needed for Agent reinforcement learning.

The open-source community is about to take off.

Today, I am willing to refer to Kimi and DeepSeek as the two saints of open source.

I also read the license for Kimi K3 this time.

Ordinary researchers, developers, and companies can use, copy, modify, deploy, fine-tune, and distribute it for free, and they can also create derivative products and commercialize them.

It only leaves two thresholds for ultra-large-scale commercial use.

If your model service business has a total revenue exceeding $20 million for 12 consecutive months, you need to sign a separate agreement with the Dark Side of the Moon.

If the monthly active users of a commercial product using K3 exceed 100 million, or if the monthly revenue exceeds $20 million, the product interface must prominently display Kimi K3.

Yes, it just needs to prominently display Kimi K3.

Oh my...

Some have already said that if Kimi K3 can bring them $20 million in a single month, they wouldn't just prominently display Kimi K3's brand logo.

They could tattoo Kimi's logo on their butt, or even get two tattoos.

Tonight, with the open-source release of Kimi K3, I feel like the entire event has reached its climax, with public opinion overwhelmingly in favor.

It has also made history in Hugging Face.

At this moment, the attitude of the United States and the open-sourcing of Kimi K3 have created a magical chemical reaction.

In the future, this event may be historically referred to as:

The Sarajevo Incident of the AI Industry.

This world needs open source.

Old Huang and his team embrace open source, and of course, beyond ideals, there is also a strong scent of money.

In 2002, Joel Spolsky wrote a classic business principle called Commoditize Your Complements, turning your complements into commodities. The cheaper the complements of something, the greater the demand for that thing.

NVIDIA sells GPUs.

The more open-source models there are, the more companies around the world are willing to deploy, fine-tune, and train themselves, ultimately requiring more GPUs.

AMD, Dell, cloud computing companies, inference service providers, the logic is similar.

Development tools and application platforms like GitHub, Vercel, Replit, LangChain, and OpenClaw also naturally hope for more, cheaper, and easier-to-replace underlying models.

Because for every dollar the model cost decreases, the application layer gains an extra dollar of space.

Old Huang firmly believes in open source, and he is one of the biggest beneficiaries of open-source models; these two things are not in conflict. But I firmly believe in and embrace open source. The reason is simple: I am an ordinary person. In my lifetime, I may never be able to train a model like K3; I don't have thousands of cards, and I know nothing. The formulas in technical reports are something I can barely understand with the help of AI. But that doesn't stop me from standing on the work, frameworks, and models that countless predecessors have open-sourced to create my own work today. Ordinary people can participate in every technological revolution thanks to those predecessors who are willing to leave the endpoint for those who come later as a starting point. We haven't suddenly become geniuses.

We just don't have to start from scratch again.

The true power of open source lies here.

What a team spends enormous computing power and countless big shots endure for a long time to obtain, once made public, will become a public starting point for countless ordinary people around the world the next day.

I believe this is the compound interest of civilization.

Finally, I extend my highest respect to all the great figures in the human world, the open-source world, and the AI world.

Thank you all.