In July, Moonshot AI released the world’s largest free artificial intelligence (AI) model. Eight weeks later, the company is filing police reports.
Kimi K3 arrived on July 16. Anyone could download it and run it. In an industry that guards its best work, that was rare. But then the servers buckled.
A Launch That Ate Its Own Computers
Four days in, Moonshot stopped selling new subscriptions. Paying members kept their seats as everyone else waited.
“Over the past 48 hours, demand has pushed close to the limits of our current capacity,” the team said in a statement.
That was the loudest the company would be all summer.
On September 11, Moonshot pushed a smaller model, K2.8 Preview, across its coding platform. Read the changelog closely and something odd appears. Some K3 requests now get answered by the smaller model instead.
However, that same week, Anthropic accused Moonshot of quietly passing Kimi users’ questions to Claude. It said Moonshot showed Claude’s answers as Kimi’s own work. In a single 10-day window, it counted nearly 300,000 handoffs routed through 5,380 fake accounts.
One of those users was probably military, Anthropic says. They uploaded footage from hundreds of security cameras in Chengdu and asked whether a man was behaving strangely.
Here is the uncomfortable part. All of it rests on Anthropic’s own logs. No regulator has tested any of it. BeInCrypto reported on September 1 that Anthropic had already shipped a block against this. That was 10 days before its September threat report named a single company.
Then came the claim that Moonshot’s founder had been taken away after the Anthropic report.
On Saturday, the company answered.
“The information circulating online regarding the founder and employees is completely fabricated and malicious slander. We have immediately reported the matter to the police and will pursue legal action against those responsible for spreading the rumors,” Moonshot AI said on Saturday.
No newspaper has confirmed an arrest. No agency has confirmed an investigation. What is confirmed is stranger. The year’s biggest open AI model is now partly handled by something smaller, and the fullest account of where its users’ questions went was written in San Francisco.