The best open model in the world is now Chinese, and it comes with conditions
Moonshot AI released the weights for Kimi K3, a 2.8 trillion parameter model that sits a few places behind the best American systems. For European companies it opens a real option, with a license and a hardware bill attached.
On July 16 Moonshot AI, the Beijing lab behind the Kimi chatbot, announced Kimi K3, and on July 27 it put the weights up for download. Anyone with enough hardware can now run one of the strongest AI models in the world on their own machines. Nathan Lambert, who writes the Interconnects newsletter on open models, called it the strongest open model ever released. On the Artificial Analysis intelligence index it ranks behind only a couple of closed American models.
What Moonshot shipped
K3 has 2.8 trillion parameters, of which 104 billion are active for any given token, and a context window of just over a million tokens. The download is 1.56 terabytes. Running it takes roughly 1.4 terabytes of memory, even at the compressed MXFP4 precision Moonshot recommends, which means a rack of GPUs.
The license is Moonshot’s own, not MIT or Apache 2.0 like most open models. Internal use is unrestricted. If you sell K3 as a hosted service and that business passes $20 million in revenue over twelve months, you need a separate agreement with Moonshot, and products with more than 100 million monthly users or $20 million in monthly revenue must display the Kimi K3 name. In practice you can do almost anything with the weights except compete with Moonshot’s own hosting business.
Moonshot’s API charges $3 per million input tokens and $15 per million output tokens for those who would rather not host it. For comparison, Anthropic launched Claude Opus 5 on July 24 at $5 and $25, pitched as close to its top model Fable 5 at half the price.
Why a Chinese lab gives this away
Moonshot’s founder Yang Zhilin has said the aim is to grow faster than closed US competitors by being more available than they are. Lambert reads K3 as evidence that Chinese labs can now match the American frontier through their own engineering, without relying on distilling US models, and notes that open-source AI has explicit backing from the top of the Chinese government.
For the closed labs this is bad news. Every capable open model caps what OpenAI, Anthropic and Google can charge for comparable work, and K3 shrinks the lead they need to justify their prices. Lambert argues that this squeezes the margins that pay for frontier research, while speeding up adoption everywhere else.
The European case for running it yourself
Open weights solve a problem closed models cannot. If you host the model yourself, your prompts and documents stay on infrastructure you control. You can run it in a Danish or German data center, inside your own security perimeter, and answer a procurement team’s data-residency question in one sentence. In June, Washington switched off Fable 5 for every non-American user. A model on your own servers keeps running whatever a foreign government decides.
Many European compliance teams will object to a Chinese model on principle, and some questions are fair. Self-hosting settles the data question, since nothing goes back to Moonshot. It says nothing about what the model was trained on or how it answers politically sensitive questions. You can test both on your own material, and should, before it goes anywhere near a customer.
Cost is the real obstacle. Hardware that holds 1.4 terabytes of model in memory costs millions of kroner to buy, or a large monthly cloud bill to rent. For most Danish SMEs, self-hosting K3 does not add up today. The realistic options are a European hosting provider that runs open models in EU data centers, or a smaller open model that is good enough for the job.
What to do with this
If you buy AI from a single closed provider, use K3 in the next price negotiation even if you never download it. Your vendor now prices against a capable free alternative, and a volume commitment signed before July is worth reopening.
If you handle sensitive data and have held back on AI because of where the data would go, open models at this level remove that objection. Evaluate on your own tasks and documents, and compare quality per krone with the closed models you would otherwise use.
Read model licenses like any other contract. K3’s weights are free to use until your business is large enough to compete with Moonshot, and at that point you need Moonshot’s permission.


