China Built a 2.8 Trillion Parameter Model Without the Chips — Kimi K3 & The Sovereignty Shock

What actually happened

Moonshot AI released Kimi K3 on July 16, 2026 — 2.8 trillion parameters, the largest open-weight model anyone has shipped. Artificial Analysis placed it third on its Intelligence Index, level with Opus 4.8 and GPT-5.5. On Frontend Code Arena it beat Claude Fable 5 outright, at roughly a third of the token price. Markets reacted the way they reacted to DeepSeek.

The detail that matters more than the benchmarks: Moonshot did this under export controls that barred it from the hardware American labs train on. Its president, Yutong Zhang, put it plainly at the World Economic Forum this year — they knew they didn’t have the luxury to simply scale up compute. So they changed the architecture instead. Eight hundred ninety-six experts, sixteen active per token, and a stack of routing and attention changes that Moonshot claims buys about 2.5x better scaling efficiency than Kimi K2.

Export controls didn’t stop the model. They changed the engineering problem.

The part the benchmark coverage misses

Mixture-of-experts sparsity is a compute strategy. It is not a memory strategy.

Sixteen experts fire per token — but the router can call any of the 896 on the next one. Every parameter has to sit in fast memory the whole time. At 4-bit quantisation, 2.8 trillion parameters is about 1.4 terabytes, before you allocate anything to the key-value cache a million-token context demands.

So the weights are open and the download is free. Holding them is a datacenter procurement decision. Open weights at this scale are open the way a shipyard is open: the blueprints are public, and the barrier moved somewhere else entirely.

What it costs

I spent 25 years in cybersecurity watching governments try to control information the way they control shipments. In the 1990s the US classified strong encryption as a munition. The mathematics travelled anyway, and by 2000 the controls had been relaxed because they were restraining American industry more than anyone else.

There is a second cost, and boards will inherit it. “Open weights” sounds like transparency. You receive the numbers — not the training data, not the provenance, and no human-readable account of what any part of the model encodes. Possession is not explanation.

The organisations that already know where those weights would physically run will move on this in weeks. The ones who wait will spend 2027 negotiating with a vendor while telling the board they had an alternative all along.