Kimi K3 is available for unrestricted, private use in the US and is nearly as powerful as the frontier models available to the public from OpenAI and Anthropic. It's more powerful than frontier models from other providers. All of these closed models have guardrails that are meant to stop bad actors from doing malicious things. Kimi K3 has no such guardrails.
K3 was made by a Chinese company but today that doesn't matter. Anyone in the world can host it anywhere they choose and use it to do whatever they like. Anyone can download the model itself, inspect it and retrain it as they see fit.
Open source is a weapon traditionally used by the smart “little guys” to challenge the incumbents in their space. In the past developing with open source has sometimes been too costly. AI tools have changed that reality.
Today K3 is the strongest defensive weapon that most small and mid-market US businesses have to protect themselves.
The future is open source!
AI is quickly becoming a commodity.
OpenAI and Anthropic are scared to death. I see their fear by how they are lashing out. Since they are the incumbents, both labs envision a dystopia when they talk about open source. Dean Ball, Head of Strategic Futures at OpenAI, writes:
“One probable outcome of an open-weight-model-dominant world is full AI communism, which is precisely what China proposes: rather than a market product, AI is a 'public good' which will ultimately be provided by the state as a kind of 'digital public infrastructure.' This future strikes me as a dystopian hellscape.”
On June 17th, well before Moonshot announced Kimi K3, Emily Chang asks Anthropic CEO Dario Amodei, “Strong open source models are coming out of China and you have US companies building on them for free. Is that a threat?”. He begins his defense stating that people don't really want to use models that are less intelligent and compares them to tech from 2023 and 2024, stating how incredibly far Anthropic is. He feels that people who have the mindset that allows them to build on open source models have instincts that are wrong. He states his concern that in 12 months when open source models have Mythos class capabilities, we will be in trouble if we haven't patched everything already. He goes on to talk about that being a dystopian time like Orwell's novel 1984.
The full exchange is here:
His 12 months lasted 30 days.
Clearly he didn't know what Moonshot had planned!
Redis is an extremely fast database that lives in memory rather than on a disk and is used in critical operations all around the world. On July 22nd, Chaofan Shou of Yeta Labs gave Kimi a goal and only 38 words to find an unknown flaw in Redis. It took 27 minutes. The next day on July 23rd he spent an hour and a half and found 19 more! That same day he found an unknown flaw that could execute any command through the Telegram Desktop or iOS app. He ended his workweek Friday by saying
“i dont think anything is safe for now. it also found real vulnerabilities in latest chrome and wechat. gl everyone”
That's one guy, publicly announcing what he's doing.
Bad actors have been using Kimi on Moonshot's server in China for almost 2 weeks now. I assume China knows what's happening on Moonshot's servers since they don't have much for privacy laws. The US is kinda sorta frenemies with China so they're not going to be openly hostile towards us. They don't need to be.
Think about what can happen when those who openly oppose western freedom and culture get ahold of this on their own servers.
Electric capacity aside, the rest of the world can use Kimi K3 to either fight or defend in equal ways.
US companies OpenAI and Anthropic have a few months edge over K3. But all their tools have guardrails for the general public. K3 can be used to fight dirty and only a small fraction of US companies have access to closed models that the public doesn't to help protect them.
Last week, OpenAI disclosed that GPT 5.6 Sol and an unnamed agent had hacked Hugging Face's systems. The only tool Hugging Face could use to defend itself was a far less powerful model, the open source GLM 5.2. Hugging Face is one of the groups probably best capable to defend in such attacks, which is why they were able to do so with GLM 5.2. A couple days later, OpenAI was hit with an outage with ChatGPT, Codex, and APIs going down together. What happened hasn't been disclosed, but I have to wonder if Kimi K3 was used. I was unable to log into my local bank just yesterday. Apparently, my account was locked down for too many failed password attempts. That definitely was not me. But then all of a sudden, today, it's fine like nothing happened. So now I'm wondering if this was a much larger, undisclosed issue that they had.
These stories are going to become commonplace.
More than ever, all individuals and businesses need to take security very seriously. Make sure to enable automatic updates on any system you can and prioritize updating anything else that does not update automatically. Separately create a secret word or passphrase for your team or your family and ideally, only share that in person or worst case, over the phone. Don't email it, don't text it, don't put it in a document.
If you are contacted by anyone you know and something seems fishy, just ask what the secret word is. The technology has already been out there before K3 was launched that can make a pretty darn convincing live video call, created with simply a short clip of video where you're talking. Even easier is an audio clone. Perhaps you've left a voicemail somewhere and that's been compromised. Now someone can pretend to be you.
All this stuff sounds pretty scary. But don't get me wrong, I love Kimi K3. And right now, for all intents and purposes, I own Kimi K3, can train it however I like and will use it for efficient and exciting, positive things. Your business can do the same.
Guidance Systems can help you own the future of your business rather than having to rely on renting a black box you can't control.