The Sequence Radar #897: Last Week in AI: China, Compression and the Open-Model Race

The last seven days made it clear that the AI race has quietly split into several races at once, and none of them are about raw benchmark scores anymore.

Thinking Machines Lab put out Inkling, a mixture-of-experts model with 975 billion total parameters but only 41 billion active at a time, handling text, image and audio with a million-token context window. It ships with open weights under an Apache 2.0 license. What stands out isn't the size — it's the framing. The company is betting people will want to shape and calibrate reasoning effort themselves rather than just rent intelligence from a single provider.

Moonshot AI took the scale question and ran further with it. Kimi K3 is a 2.8-trillion-parameter model that only activates 16 of its 896 experts at once, also supports a million-token context and vision, and is built for long, complex coding and knowledge work. Moonshot calls it the first open model in the three-trillion-parameter class, though the full weights won't actually land until later this month. Worth flagging that gap between announcement and availability — but the bigger signal is that Chinese labs are now treating openness itself as a competitive weapon, not just a cost play.

Then there's PrismML's Bonsai 27B, which goes the opposite direction entirely: down. Its one-bit version is just 3.9GB, small enough, the company claims, to run inside a normal smartphone's memory, while still holding onto most of the original model's performance on its benchmarks. Nobody's independently verified that yet. But if it holds up, compression this aggressive could matter as much as raw scale — it changes who can run capable AI, and where.

OpenAI's contribution this week wasn't a chatbot at all. GPT-Red is an internal system trained through self-play to attack other models and find their weaknesses. In testing on unfamiliar scenarios, it broke through GPT-5.1's defenses 84% of the time, versus 13% for human red-teamers, and its discovered attacks were then used to harden GPT-5.6. That's a different kind of scaling — not bigger models, but attackers and defenders escalating against each other at machine speed.

And all of this landed the same week Xi Jinping stood in Shanghai and pitched open-source AI as a global public good, pushing a new international cooperation body while offering Chinese AI support to developing countries. That's not just rhetoric floating above the technology. Standards and ecosystems create dependencies just as effectively as hardware does, which means the fight over who builds the smartest model is increasingly tangled up with the fight over who everyone else ends up relying on.