Kimi K3 is now yours to download. Good luck running it
Moonshot dropped the full weights of Kimi K3 on Hugging Face today, and it's the strongest open-weight model anyone has ever released. 2.8 trillion parameters, mixture-of-experts with about 50 billion active per token, a million-token context, native multimodality. On the Artificial Analysis index it lands at #3, behind only Claude Fable 5 and GPT-5.6 Sol. It's #1 on Frontend Code Arena. And the license is a modified MIT, so you can build a business on it. Training data and code stay closed, but the weights are yours.
Then you read the spec sheet and the mood changes. Even squeezed down to four-bit MXFP4, the model is a 1.4 terabyte file. Moonshot's own guidance says plan for at least 64 accelerators to serve it. So yes, you can now run a frontier-class model with no per-token bill and nobody watching your prompts. You just need a small data center to do it. Nathan Lambert called it the open-weights escalation, and someone else nailed the real headline: open weights, closed wallet. The freedom is real, the barrier just moved from the API key to the electricity bill.
The timing is the loud part. The weights land the same week Xi spoke at the World AI Conference in Shanghai, the Nasdaq slid about 1% on Friday, and Nvidia sold off as investors did the math on what free frontier models do to everyone's margins. That's the actual story here, not benchmarks. Strong open weights don't just pressure OpenAI and Anthropic's pricing, they threaten the entire premise that a frontier model is a scarce product you rent. Washington is now openly debating whether to restrict models like this, which is its own kind of absurd. You cannot ban a file that's already mirrored on a thousand machines.
So the question K3 forces isn't which model tops which leaderboard. It's whether an open frontier model is infrastructure or contraband. Moonshot just made everyone answer it. Weights are here: https://huggingface.co/moonshotai
← Back to all articles
Then you read the spec sheet and the mood changes. Even squeezed down to four-bit MXFP4, the model is a 1.4 terabyte file. Moonshot's own guidance says plan for at least 64 accelerators to serve it. So yes, you can now run a frontier-class model with no per-token bill and nobody watching your prompts. You just need a small data center to do it. Nathan Lambert called it the open-weights escalation, and someone else nailed the real headline: open weights, closed wallet. The freedom is real, the barrier just moved from the API key to the electricity bill.
The timing is the loud part. The weights land the same week Xi spoke at the World AI Conference in Shanghai, the Nasdaq slid about 1% on Friday, and Nvidia sold off as investors did the math on what free frontier models do to everyone's margins. That's the actual story here, not benchmarks. Strong open weights don't just pressure OpenAI and Anthropic's pricing, they threaten the entire premise that a frontier model is a scarce product you rent. Washington is now openly debating whether to restrict models like this, which is its own kind of absurd. You cannot ban a file that's already mirrored on a thousand machines.
So the question K3 forces isn't which model tops which leaderboard. It's whether an open frontier model is infrastructure or contraband. Moonshot just made everyone answer it. Weights are here: https://huggingface.co/moonshotai
Comments