Kimi K3.1 Pro weights arrive as a 1.56 TB download

Moonshot AI released downloadable weights for Kimi K3.1 Pro on November 4, opening a self-hosting route for organizations with suitable servers. The release fulfills the download promised when the company introduced the model through its apps and API last month.

The official Hugging Face repository is publicly readable without an access approval. Its file listing contains 96 safetensors weight shards totaling approximately 1.56 TB, alongside configuration, tokenizer, processing code and a license. AI Ping checked the listing and file sizes; the complete checkpoint was outside this verification.

Moonshot's model card lists 2.8 trillion total parameters, with 104 billion active per token. The model accepts text and images and supports a 1,048,576-token context. The company describes quantization-aware training using MXFP4 weights and MXFP8 activations, the intermediate values used during computation. The card also describes replacing K3’s full-attention layers with sparse attention in 24 layers, to make long-context processing more efficient by selecting relevant positions.

The download is a separate milestone from October's hosted launch. Moonshot described K3.1 Pro as a post-trained upgrade of K3 for coding, 3D creation and financial analysis, and promised full weights by November 4. It also introduced a faster Highspeed service. The checkpoint directly verified for this report is the K3.1 Pro repository.

October's technical post described preparation of the weights. It said it removed cyber-related material from mid-training data and reinforcement-learning environments. On its internal vulnerability benchmark, the final model scored 81.0% at pass@3, versus 90.5% for a checkpoint without those controls. These are company-reported measurements.

Self-hosting starts at server scale

The vLLM deployment guide linked by the model card gives an eight-GPU NVIDIA baseline: one B300 node or two GB300 trays. The guide calls this its simplest working NVIDIA setup and says throughput was not tuned.

The AMD MI355X entry qualifies a specific configuration for 48 concurrent requests. Its reported execution is limited to that configuration.

SGLang's separate cookbook marks command configurations as verified, pending verification or unverified. It reports verification for selected B300 decode configurations while other combinations remain in progress. The unit being qualified is an exact hardware-and-command combination.

For coding agents and other multi-turn workloads, the cookbook explains why reusing repeated prompt prefixes matters. Its deployment options cover moving conversation caches beyond GPU memory and keeping a session on workers that already hold its prefix. Those workload choices sit alongside the hardware qualification.

The accompanying custom Kimi K3.1-Pro License permits use, modification, fine-tuning and redistribution, subject to conditions including retention of notices. A model-as-a-service business whose combined revenue with affiliates exceeds $20 million over any consecutive 12 months needs a separate Moonshot agreement before commercial use.

Commercial products exceeding 100 million monthly active users or $20 million monthly revenue must display the model name prominently. These two additional requirements exempt internal use that exposes no model, outputs or capabilities to third parties, and use through Moonshot's official products or certified inference partners. The complete terms accompany the files.