China's Coding Model Surge and the SpaceX Cursor Acquisition | 18th June 2026

**SPEAKER_1** (0:00)
Welcome to Colaberry AI Podcast, brought to you by Colaberry AI Research Labs and Carl Foundation. Right now, there is basically an invisible war happening for the future of software.

**SPEAKER_2** (0:12)
Oh, absolutely.

**SPEAKER_1** (0:12)
Yeah. On one side, you have developers in China who are releasing these massive trillion-parameter coding brains to the public for practically free, and then on the other side, tech billionaires are dropping like $60 billion just to watch how you type.

**SPEAKER_2** (0:28)
The datagram.

**SPEAKER_1** (0:29)
Exactly. Today, we are taking a deep dive into this massive, highly technical tectonic shift in the AI landscape. I mean, it is not just a single new model dropping today.

**SPEAKER_2** (0:40)
A whole wave.

**SPEAKER_1** (0:41)
It's a flurry. Yeah. We're talking highly efficient, open weight coding models, a staggering acquisition in the developer space, and this fundamental architectural shift in how voice AI actually processes sound.

**SPEAKER_2** (0:53)
Which is huge.

**SPEAKER_1** (0:53)
It really is. Now, I want to warn you, right up front, we are going to get deep into technical weeds today.

**SPEAKER_2** (0:58)
We really are. Yeah.

**SPEAKER_1** (0:59)
We are talking parameter counts, speculative decoding mechanisms, mixture of experts architecture, and specific benchmark validations. But I promise, we will make it all make sense. Okay. Let's unpack this.

**SPEAKER_2** (1:12)
It is an incredibly dense stack of technical developments to go through today. To really understand why these new open weight models are making such a sudden massive impact, we have to start with the geopolitical catalysts that kind of forced the market's hand here.

**SPEAKER_1** (1:28)
Right. The context is everything.

**SPEAKER_2** (1:30)
Exactly. If we impartially look at the recent timeline regarding Anthropics Clawed, the context becomes really clear. Following the Trump administration's expert control directive, which reportedly blocked foreign nationals from using the service, highly capable models like Fable 5 and Mythos 5 were taken offline completely.

**SPEAKER_1** (1:48)
Wow. Just completely dark.

**SPEAKER_2** (1:50)
Yeah. Offline for everyone. And as an impartial observer of the tech industry, the critical takeaway from that event isn't the politics at all, it is the sheer enterprise panic it caused.

**SPEAKER_1** (1:59)
Oh, I mean, that makes total sense. Because if you are a CTO and you have built your entire company's automated software pipeline around a specific API, hitting a sudden brick wall like that is an enterprise nightmare.

**SPEAKER_2** (2:12)
It really is.

**SPEAKER_1** (2:13)
It raises this terrifying question. What happens to your business if your access to a closed proprietary model just disappears overnight?

**SPEAKER_2** (2:22)
That unreliability, that exact fear, is what triggered an immediate aggressive market response. Enter ZAI, which is the lab formerly known as Jepu AI.

**SPEAKER_1** (2:31)
Oh, right, the rebrand.

**SPEAKER_2** (2:32)
Yeah. On June 16th, they released Glm 5.2. This wasn't just a minor update.
They dropped a 753 billion parameter open weights model, specifically engineered for long-horizon coding agents.

**SPEAKER_1** (2:46)
753 billion, that is massive.

**SPEAKER_2** (2:48)
Massive. But the critical factor here is the licensing. It was released under a pure, open, completely unrestricted MIT license with zero regional limits.

**SPEAKER_1** (2:57)
So anyone can use it?

**SPEAKER_2** (2:58)
Anyone. They are explicitly calling it technical access without borders.

**SPEAKER_1** (3:02)
So they are directly answering that enterprise panic by saying, hey, here are the weights, host it yourself, control your own destiny. But a 753 billion parameter model is a beast.

**SPEAKER_2** (3:14)
It's heavy, yeah.

**SPEAKER_1** (3:15)
The hardware requirements alone must be staggering, especially because it boasts a 1 million token context window. I mean, how do you even load a million tokens of code without the memory just completely crashing?

**SPEAKER_2** (3:27)
Well, that is the core engineering hurdle, right? In standard transformer architectures, maintaining a 1 million token context requires an astronomical amount of memory to hold what is called the key value cache or KV cache.

**SPEAKER_1** (3:41)
Okay, the KV cache.

**SPEAKER_2** (3:42)
Yeah. Think of the KV cache as the model's scratch pad for short-term memory.

**SPEAKER_1** (3:46)
Like a working memory space.

**SPEAKER_2** (3:48)
Exactly. Every time it reads a word, it saves the mathematical representation of that word, so it doesn't have to recalculate it later.

**SPEAKER_1** (3:53)
Oh, that makes sense.

**SPEAKER_2** (3:54)
But when you hit a million tokens, that scratch pad gets impossibly large. So to solve this, Glm 5.2 utilizes a novel method called index share.

**SPEAKER_1** (4:03)
Index share, okay.

**SPEAKER_2** (4:05)
Yeah. Before we get to the sharing part, we have to note that they use sparse attention layers.

**SPEAKER_1** (4:10)
Right.

**SPEAKER_2** (4:10)
Instead of every token looking at every other token, which scales quadratically and just destroys compute sparse attention, forces the model to only pay attention to the most mathematically relevant tokens.