Kimi K3: open model with 2.8 trillion parameters
Moonshot AI has released the Kimi K3 weights on Hugging Face. It is the largest model to date whose weights can be downloaded freely, and it ranks near the top of independent leaderboards.
- Moonshot AI published the Kimi K3 weights on July 27. At 2.8 trillion total parameters it is the largest model with publicly available weights.
- The architecture is Mixture-of-Experts: roughly 104 billion parameters are active per token, or 16 of 896 experts.
- The context window is one million tokens, text and image input is native, and the Hugging Face repository weighs about 1.56 terabytes.
- The model ranks third on the Artificial Analysis index, second on Vals AI and first in the Frontend Code Arena.
The Chinese company Moonshot AI released the weights of its new Kimi K3 model on Hugging Face on July 27, a day after the Sunday for which the release had been announced. With 2.8 trillion total parameters it is the largest model whose weights can be downloaded freely, and on independent leaderboards it sits immediately behind the strongest closed systems built by American companies.
What makes the release more interesting than the headline number is where the model lands. On the intelligence index maintained by Artificial Analysis, Kimi K3 is third, behind Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol Max. It is second on the Vals AI index and first in the Frontend Code Arena, which measures generation of user interface code. Until now open models have as a rule trailed the commercial frontier by several generations.
What sits behind the 2.8 trillion figure
The total parameter count does not mean all of it runs at once. Kimi K3 uses a Mixture-of-Experts architecture in which only part of the network is activated for each token: according to the official model card, around 104 billion parameters, or 16 of the 896 experts. The model accepts a context of 1,048,576 tokens, slightly over one million, and handles text and images natively through its own MoonViT-V2 vision encoder of 401 million parameters.
The long context rests on a new attention mechanism called Kimi Delta Attention, combined with a gated variant of multi-head latent attention. The company says the new architecture delivers about 2.5 times better overall scaling efficiency than the previous Kimi K2 model. That claim comes from the vendor and has not yet been independently verified.
The weights are published under the company’s own terms, called the Kimi K3 License. Training data and training code were not released, so it is more precise to call this an open weights model rather than open source. Through the Kimi platform the company also offers access via an interface compatible with the OpenAI and Anthropic standards, which in practice means existing integrations can be switched over with few changes.
Are open models really catching up with closed ones?
The analyst Nathan Lambert estimates that the gap between open and closed models, until recently discussed as six to nine months, has now narrowed to three to five months. That is an estimate rather than a measurement, and it leans on leaderboards that cover narrow sets of tasks.
It is worth saying what the release does not prove. A score on an intelligence index is not evidence that the model performs better in a specific production setting, because leaderboards measure a limited set of tasks and can be optimised for. First place in the Frontend Code Arena speaks to interface code generation, not to reliability in legal, medical or financial use. Independent checks of efficiency and quality are still to come.
The practical limit is physical. The Hugging Face repository weighs about 1.56 terabytes, and running a model of this size requires a GPU cluster that few outside large companies and research centres possess. A free download does not mean free use on a home computer. What does change is who is allowed to test, fine-tune and run the model without an intermediary, and that difference reaches into the question of data sovereignty.
Do open weights represent a real alternative to commercial services for you, or is the cost of the infrastructure still too high?
