China's Moonshot AI releases Kimi K3.1 Pro weights: 2.8 trillion parameters, an 8-GPU server estimated at $550,000-$650,000 to run

China's Moonshot AI published the weights of Kimi K3.1 Pro on Hugging Face on Nov. 4. It is a mixture-of-experts model with 2.8 trillion total parameters, 104 billion active per token.

OrcaRouter's analysis found the repository holds 111 files totaling 1.56 TB. It ships in the MXFP4 precision it was trained to serve in, far below the 5.6 TB BF16 figure widely quoted. Context is about 1.05 million tokens; it takes text and images. Moonshot describes it as heavier reinforcement learning on the K3 base rather than a new architecture. AI commentator Chubby (kimmonismus) pointed to a server of 8 Nvidia B300 GPUs (2.3 TB of memory) at an estimated $550,000-$650,000. The license is not a standard open-source one: operators with more than $20 million in annual revenue who offer inference or fine-tuning to third parties need a separate agreement, and products above 100 million monthly users or $20 million in monthly revenue must display "Kimi K3.1-Pro."

Yuchen Jin said on X that K3.1 Pro launches on Databricks on day 0, calling it "the best oss model to date" that his team already uses for internal coding. vLLM announced day-0 support.

Reactions split. One user quipped "open weights, closed wallets." Creator Alex Finn claimed Qwen 3.8 on an RTX Spark laptop gives "Opus 4.6 performance, but unlimited, private, and local" (about 1,400 likes; his own assessment, not independently verified).