AI in Motion: Moonshot's 2.8T-Parameter Model Redefines Open Source Frontier

China just released an open-source model with 2.8 TRILLION parameters. In our last episode, we talked about the Inkling model launch as a real shift. Today, the scale of what's possible just jumped by an order of magnitude. Moonshot AI's Kimi K3 isn't just another model—it's a statement. It’s the largest open-source-track AI model ever created, and it changes the calculus for everyone. Here’s what else is moving. First, Google finally launched Gemini 3.5 Pro. It arrives with a massive two-million-token context window and a new reasoning mode called "Deep Think." But it’s also six weeks late, and that delay is a story in itself. Second, the money keeps flowing, but not just where you'd expect.

AI infrastructure startup Fireworks AI just closed a one-point-five BILLION dollar Series D round. Their valuation is now seventeen-point-five billion dollars. They're processing over forty trillion tokens a day—that’s the engine room of this entire industry getting a massive upgrade. Third, a smaller but critical signal. An AI startup called Sable raised forty-five million dollars from Sequoia to build "AI employees." They’re less than a year old and already in production at companies like Notion. This isn't a research project; it's about automating complex customer interaction right now. And finally, all of this is happening as the World Artificial Intelligence Conference kicks off in Shanghai.

Over three hundred new AI products are on display. This isn't just a trade show; it's proof of a new, multipolar AI capital landscape. The U.S. monopoly on funding innovation is breaking. Okay, let's go deeper on the two biggest stories: Google's comeback and China's power play. They look like separate events, but they are two sides of the same coin. Let's start with Google. Gemini 3.5 Pro is here, and the two-million-token context window is a serious capability. The new "Deep Think" mode, available on their two-hundred-fifty-dollar Ultra tier, is their direct shot at taking back the reasoning crown. But here's the catch. They are late. The reason for the six-week delay is that Google's engineers found structural failures in the original base model.

So they scrapped it. They started pretraining all over again. That is a HUGE and expensive decision. It signals that Google refused to ship a flawed flagship. That's the responsible engineering call, but it gave their rivals an opening. OpenAI’s GPT-5.6 and Grok 4.5 launched just days earlier with aggressive pricing. So now, for the first time in a while, Google is playing catch-up. They have to fight on features and on price, and they're doing it from behind. Now, contrast that with Moonshot AI in Beijing. Their Kimi K3 model is a completely different strategy. At 2.8 trillion parameters, it's a sparse Mixture-of-Experts model—that's how they achieve that scale. But the most important detail isn't the size.

It's the plan. They are releasing the full open-source weights on July 27th. This isn't just an API you can rent. This is a foundational tool they are giving away to the entire world. They are arming the open-source community with a weapon that, until yesterday, was reserved for a handful of trillion-dollar corporations. And investors are betting on this strategy. Moonshot AI is now valued at thirty-one-point-five billion dollars. While that's still a fraction of its U.S. peers, it's a clear signal that the game is changing. What you're seeing today is a fundamental split in strategy. In the U.S., a titan chose quality over speed and is now fighting to regain its footing. In China, a challenger is choosing scale and openness to try and rewrite the rules of the entire market.

The most exciting thing in AI right now isn't just another frontier model. It's about who gets to build the software of the future. The answer used to be "whoever can afford to rent it from Silicon Valley." That is no longer the only answer. The race is now multipolar, and the starting gun just fired in Beijing.