Kimi K3.1 Pro's open weights: 1.56 TB, MXFP4, and a $20M revenue license line (theainewsreport.com)

Moonshot has launched the Kimi K3.1 Pro, a cutting-edge AI model featuring 2.8 trillion parameters, with its open weights now available for public download on Hugging Face. The total file size is approximately 1.56 TB, stored in 96 safetensors files utilizing the MXFP4 format, which averages out to about 4.5 bits per parameter. Notably, this mixture-of-experts model employs 104 billion active parameters per token, requires over 1.5 TB of fast memory to operate, and is designed for use on multi-GPU servers. The model achieved a score of 65.7% on Terminal-Bench 4.0, positioning it competitively against similar models like Claude Opus 5.5 and GPT-6 Astra. Significantly, the licensing model for Kimi K3.1 Pro introduces unique stipulations: companies generating over $20 million in revenue must negotiate a separate agreement for commercial use, while internal use remains exempt. This move could influence how organizations approach AI model deployment and monetization. The Kimi K3.1 Pro also includes a powerful integrated vision encoder capable of processing both text and images, and its extensive parameter count optimizes performance at scale. As developers and businesses explore its capabilities, the need for strategic planning around storage, bandwidth, and GPU architecture will be paramount to fully leverage this advanced model.