• Sources: primary, discussion
  • Thread: Kimi K3.1 Pro open weights
  • Summary: The Hugging Face model card lists 2.8T total and 104B activated parameters, 896 experts with 16 selected per token, a 1,048,576-token context, and MXFP4 weights with MXFP8 activations from quantization-aware training. It is released under a custom Kimi K3.1 Pro License, and Moonshot recommends vLLM, SGLang or TokenSpeed for serving. Multi-turn and tool calls require passing back the full assistant message, including reasoning_content.
  • Why it matters: A frontier-class open-weight agentic model can now be self-hosted, though its size needs multi-node GPU capacity and its license needs a legal read.