Moonshot posted the Kimi K3.1 Pro weights on HuggingFace as promised. This is the model according to Moonshot, which almost ran into Opus 5.5 in Terminal-Bench 4.0: MoE, 104 billion active out of 2.8 trillion dollars, in the context of 1 million tokens. The weights are correct in MXFP4, the model was trained with quantization in mind, but the buffer still weighs 1.56 TB. It is recommended to run on vLLM, SGLang or kenSpeed, recipe links for all three are on the model card. Compared to K3, the layers of full attention have been replaced with safari layers, such as DeepSeek, to make the model work more effectively in a long context. The license allows commercial and fine-tuned use. A separate agreement with Moonshot is required by companies that give third-party customers access to a service model, for example through an API, if the company's total revenue is more than $20 million over 12 months. Channel @neuro.

Moonshot posted the Kimi K3.1 Pro weights on HuggingFace as promised. This is the model according to Moonshot, which almost ran into Opus 5.5 in Terminal-Bench 4.0: MoE, 104 billion active out of 2.8 trillion dollars, in the context of 1 million tokens. The weights are correct in MXFP4, the model was trained with quantization in mind, but the buffer still weighs 1.56 TB. It is recommended to run on vLLM, SGLang or kenSpeed, recipe links for all three are on the model card. Compared to K3, the layers of full attention have been replaced with safari layers, such as DeepSeek, to make the model work more effectively in a long context. The license allows commercial and fine-tuned use. A separate agreement with Moonshot is required by companies that give third-party customers access to a service model, for example through an API, if the company's total revenue is more than $20 million over 12 months. Channel @neuro.