Keeping your data in-house just got a lot more powerful
Today the world's largest open AI model released its full weights — for the first time, teams with sensitive data can run near-frontier AI on their own servers without sending anything to anyone else.
- 2.8 trillion
- total parameters
- 1.4 TB
- storage needed (compressed)
- #2
- ranking on Vals AI Index
- ~18
- high-end server chips required
Today, a Chinese startup called Moonshot AI released the full weights of Kimi K3 — a 2.8-trillion-parameter model that ranks among the top three AI systems in the world. You can now download and run it yourself. This is genuinely new. Until today, getting that level of AI capability meant sending your data to someone else's servers.
This matters most to teams that cannot or will not send sensitive data to a third-party cloud: law firms, banks, hospitals, government agencies. When you run a model on your own servers, your documents, contracts, and client data never leave your building. Open AI models have existed before, but none this close to the frontier. A near-frontier AI is now self-hostable for the first time.
The catch is real: the model's weights take up 1.4 terabytes of fast memory, even compressed. Running it needs roughly 18 high-end server chips — the kind used in large data centres, not a startup's rack. A laptop, a small cloud instance, or a single powerful workstation cannot do it. Only organisations with serious server infrastructure can actually run it.
Still, the direction is clear. A year ago, the best AI was entirely behind closed APIs. Today Kimi K3 sits just behind Claude Fable 5 and GPT-5.6 on standard benchmarks, and leads the field for writing front-end code. The gap between the best AI you can call and the best AI you can own is closing. If your team handles sensitive data and has real server capacity, evaluate Kimi K3 for self-hosting now. For everyone else: the hardware floor is dropping — twelve months from now, running frontier AI yourself will be a lot more accessible.