What did Moonshot actually release with the Kimi K3 weights, and who can run them?

Yesterday, Moonshot AI put the weights of Kimi K3 on Hugging Face. Weights are the file of learned numbers that make up a model, so anyone can now download the model that Moonshot launched as a hosted service on July 16. In its launch post, the company had promised the weights by July 27, and it kept that date.

Moonshot calls K3 the first open model in the "3T class". It is also a practical test of what "open" means. You may download it, but very few organizations own enough hardware to run it.

What was released

The download holds the model, its configuration and its code, along with a technical report on GitHub. The main numbers from the model card:

A parameter is one of those learned numbers. K3 is a mixture-of-experts model: its feed-forward layers are divided into many sub-networks, the experts, and a small router picks a few for each token. A token is a small piece of text, often part of a word. So 104 billion is roughly the compute cost per token, while 2.8 trillion is what has to be stored.

MXFP4 is a compact 4-bit number format. Moonshot trained with this format in mind from the supervised stage onward, rather than compressing the finished model, which is why the download is much smaller than 2.8 trillion parameters would otherwise suggest.

How it works, and why memory is the barrier

Picture a library with 896 reference desks. Each question sends you to 16 of them, so your visit is quick, but the building still has to hold all 896. That is the trade-off of a sparse model: low compute for each token, high memory for the whole.

Here is the arithmetic. A 1.56 TB file on GPUs with 80 GB of memory each, a common size for NVIDIA's H100, needs about 20 cards just to hold the weights, before any working memory for the conversation. The vLLM recipe, the serving software's setup guide, lists at least eight GB300 chips or eight AMD MI355X or MI350X chips, and recommends several machines for real production traffic. The New Stack describes a realistic self-hosted setup as eight or more servers with eight H100 or B200 accelerators each. Those two descriptions differ in scale, so plan from your own traffic and test it.

Moonshot recommends three serving engines: vLLM, SGLang and TokenSpeed. The model always thinks before answering. Effort can be set to low, high or max, and max is the default. In multi-turn use and tool calls, the whole assistant message, including its reasoning, must be sent back as it came.

How good is it?

Moonshot's own table puts K3 close to the leading closed models on many tests and behind them on others. Its launch post says plainly that overall performance still trails two closed models, GPT-5.6 Sol and Claude Fable 5.

Read these as claims. Different models used different harnesses, the agent software wrapped around them, and the footnotes say some scores come from third parties. A Morningstar analyst, quoted by The New Stack, said he would hesitate to call K3 nearly equal to American frontier models on real work.

What is disputed

The release comes with an argument. The New Stack reports that Anthropic and US officials, including White House technology policy director Michael Kratsios, accuse Moonshot of distillation, which means training a model partly on the outputs of another. Anthropic's head of public policy calls that theft of intellectual property. Moonshot denies it, and a Moonshot executive attributes K3's gains to its new attention designs. Some analysts have questioned whether the time since Fable 5 appeared allows large-scale distillation. None of this has been settled in public, so treat it as an open allegation. Companies buying AI may still want to ask how it affects their compliance and procurement risk.

Can a small company use Kimi K3 without asking Moonshot?

Usually yes. The Kimi K3 License starts with broad permission to use, modify, fine-tune and sell the software, including the weights. Three conditions follow.

Notice that the revenue test in section 2 counts the whole group, not only the K3 product. A provider with $25 million in unrelated revenue is covered as soon as it runs a model-as-a-service business. Merely relaying requests to a model hosted by someone else does not count.

So this is not MIT or Apache 2.0, and it is not open source in the strict sense, because the extra conditions limit some commercial uses. For a small team, the harder question is cost. Hosted access through Moonshot or an inference provider will often make more sense until your volume justifies a cluster.