Moonshot AI released the full weights of Kimi K3.1 Pro on Hugging Face Inc. on November 4, putting a 2.8-trillion-parameter mixture-of-experts model into open circulation with conditional commercial rights1,2. The release spans 96 safetensors shards totaling roughly 1.56 terabytes.
The model activates approximately 104 billion parameters per inference and supports a context window of around one million tokens. Kimi K3.1 Pro builds on Kimi K3 with expanded reinforcement learning for coding, 3D creation, financial research and long-running agent tasks. It outperforms GPT-6 Astra and Claude Fable 5.1 on some benchmarks, according to Moonshot AI.
Free commercial use is permitted, but API providers with more than $20 million in revenue over the past 12 months must sign a separate agreement, and products with more than 100 million monthly users must display the model name.
Moonshot AI excluded attack-related content from its training data, lowering the pass rate on vulnerability-exploitation tasks from 90.5% to 81.0%.
ANALYSIS The gap between the unrestricted 90.5% pass rate and the filtered 81.0% rate shows the training-data exclusion reduced but did not eliminate the model's exploit-generation capability.
Huawei Technologies announced inference support for Kimi K3.1 Pro through vLLM Ascend and SGLang, covering its Atlas A3 and Ascend 950 series systems.
ANALYSIS Huawei's day-of-release inference support gives deployers a non-Nvidia serving path for a frontier-scale open model.
The weights release follows the model's earlier availability through Kimi's applications and API, which began in late October. Vector Wire covered the initial weight publication on November 5[1].