On November 5, the full weight of Kimi K3.1 Pro was officially revealed to the public. The supporting model card simultaneously disclosed the architectural parameters, the six-model benchmarking evaluation results, and the full text of the self-development license. The reporter noticed that the underlying architecture of K3.1 Pro continues K3: total parameters 2.8 trillion, activation parameters 104 billion, and a context window of 1 million tokens. Capacity improvement mainly comes from enhanced learning in the post-training phase. Weights are quantized using MXFP4 with low precision. 2.8 trillion parameters require about 5.6 TB to be stored with 16 bit accuracy, and quantized compressed to 1.56 TB. On the same day, Arena's Agent Arena ranking data was revealed to the public. K3.1 Pro ranked first in the open weight model, ranked 13th among all participating models, and ranked first in the open model in the three categories of work, programming, and chat. The above list is based on over 17,500 real user long-term smart sessions. The net improvement rate of K3.1 Pro was +5.1%, and the previous generation K3 was only +0.9%.

On November 5, the full weight of Kimi K3.1 Pro was officially revealed to the public. The supporting model card simultaneously disclosed the architectural parameters, the six-model benchmarking evaluation results, and the full text of the self-development license. The reporter noticed that the underlying architecture of K3.1 Pro continues K3: total parameters 2.8 trillion, activation parameters 104 billion, and a context window of 1 million tokens. Capacity improvement mainly comes from enhanced learning in the post-training phase. Weights are quantized using MXFP4 with low precision. 2.8 trillion parameters require about 5.6 TB to be stored with 16 bit accuracy, and quantized compressed to 1.56 TB. On the same day, Arena's Agent Arena ranking data was revealed to the public. K3.1 Pro ranked first in the open weight model, ranked 13th among all participating models, and ranked first in the open model in the three categories of work, programming, and chat. The above list is based on over 17,500 real user long-term smart sessions. The net improvement rate of K3.1 Pro was +5.1%, and the previous generation K3 was only +0.9%.