Kimi K3.1 Pro Open Weights; vLLM Provides an Eight-GPU Deployment Recipe

Moonshot AI has released the Kimi K3.1 Pro weights, bringing a model previously available through its products and API into the self-hosted deployment stage. The company had initially promised a release by November 4; the weights, configuration, and processing code are now available on Hugging Face, and the repository is about 1.56 TB. This lets teams test the model in their own environments, while making deployment costs and reproducibility key areas for the next round of validation. Official announcement, model files

The model uses a mixture-of-experts architecture with 2.8 trillion total parameters, activating about 104 billion parameters per token. It supports text, images, and a 1,048,576-token context window. Compared with K3, the architecture replaces the original full-attention layers with Gated DSA sparse attention, alongside Kimi Delta Attention. The model card lists a quantization-aware training configuration with MXFP4 weights and MXFP8 activations. Using fewer active parameters can reduce compute, but the full set of expert weights still creates substantial storage requirements. Model card

The vLLM recipe, updated November 5, provides an eight-GPU tensor-parallel baseline, including two GB300 trays or one B300 node, and also lists an MI355X configuration. For long contexts, it supports DCP8 distributed cache parallelism. The documentation says this offers greater capacity than standard TP8, but decoding steps for the same batch may be 7% to 25% slower. Deployment choices should therefore reflect the number of long sessions, cache requirements, and latency targets—not just the maximum context length. vLLM deployment recipe

Agent integration also has a specific requirement: multi-turn conversations and tool calls must return the complete assistant message, including reasoning_content and tool_calls. The official Terminal-Bench 4.0 score is 65.7%, but it was obtained with a specified agent framework and inference settings. Cross-model comparisons should account for test conditions, and Chinese-language engineering tasks should be retested independently. Model card and evaluation notes

The weights use a custom license. Providers operating model services above a revenue threshold must sign a separate agreement, and large products are subject to additional attribution requirements and exceptions. Engineering teams should first pin the inference version, validate tool calls, long-context accuracy, and actual throughput, then decide on service scale. Full license