Kimi K3.1 Pro Weights: 1.45 TiB, MXFP4 and No Official VRAM Table
Contains affiliate links; your price doesn't change · Disclosure · How we research
TL;DR
Kimi K3.1 Pro is a 2.8T-parameter mixture-of-experts model with 104B active parameters. Its 96 public safetensors files total 1,561,293,160,568 bytes, about 1.45 TiB. The weights use MXFP4 and activations use MXFP8. Moonshot names vLLM, SGLang and TokenSpeed as recommended inference engines. It does not give a minimum VRAM or GPU-count table. Most teams should start with the API unless they have a verified serving plan.
How this article was made: This article is research-based. Moonshot's model card, license and complete repository file list were checked on November 6, 2026 at 10:30 Taipei time. The file size is the sum of repository metadata, not a hash-verified full download. We did not load the model or test a GPU server. Display name: CommX.
Moonshot has published the full Kimi K3.1 Pro weights. The 96 safetensors files total about 1.45 TiB. The model card still gives no minimum VRAM or recommended GPU count.
Kimi K3.1 Pro is a natively multimodal mixture-of-experts model. It has 2.8T total parameters and 104B active parameters, with a 1-million-token context window.
Weight and hosting facts
| Item | Official position | Practical check |
|---|---|---|
| Weight files | 96 safetensors shards | Confirm disk space and the complete file list first |
| Safetensors size | 1,561,293,160,568 bytes, about 1.45 TiB | Leave additional room for your deployment workflow |
| Architecture | MoE, 2.8T total and 104B active parameters | Active parameter count is not a VRAM formula |
| Quantization | MXFP4 weights and MXFP8 activations, with quantization-aware training | Check engine support for this format |
| License | Kimi K3.1-Pro License | Commercial use is allowed, subject to its conditions |
| Recommended engines | vLLM, SGLang and TokenSpeed | Follow an engine's official deployment recipe |
| Minimum VRAM and GPU count | Not supplied on the official model card | Do not buy hardware from an invented card table |
The license requires a separate agreement for a Model as a Service business when aggregate revenue exceeds $20 million over any consecutive 12 months. It also requires prominent Kimi K3.1-Pro attribution when a commercial product exceeds 100 million monthly active users or $20 million in monthly revenue. The text includes exceptions for internal use and access through Moonshot or certified inference partners. This summary is not legal advice.
Start with the API or prepare a real serving plan
The API is the simpler start for an application with a deadline. The model card lists model ID kimi-k3.1-pro on Moonshot's platform. This page does not reproduce API rates because its focus is weights and hosting.
Self-hosting needs more than a large download. Confirm the engine recipe, accelerator support, storage, networking and operations plan. Moonshot has not supplied the purchasing table that would turn 1.45 TiB of files into a supported server configuration.
Check the official Kimi API platform
Read the official Kimi K3.1 Pro model card
Sources (checked November 6, 2026, Taipei)
- https://www.agentcomm.net/zh/kimi-k3-1-weights-hosting/
- https://huggingface.co/moonshotai/Kimi-K3.1-Pro
- https://huggingface.co/moonshotai/Kimi-K3.1-Pro/raw/main/LICENSE
- https://huggingface.co/api/models/moonshotai/Kimi-K3.1-Pro/tree/main?recursive=true&expand=true
FAQ
Are the complete Kimi K3.1 Pro weights public?
Yes. The official moonshotai/Kimi-K3.1-Pro repository lists 96 safetensors shards. We rechecked the complete file list on November 6, 2026.
How large is the Kimi K3.1 Pro download?
The 96 safetensors files total 1,561,293,160,568 bytes. That is about 1.45 TiB, before allowing extra working space for download and deployment.
How much VRAM does Kimi K3.1 Pro need?
Moonshot does not provide a minimum VRAM or recommended GPU-count table on the model card. This page does not estimate one. Check an official inference-engine recipe for your chosen deployment.
Can Kimi K3.1 Pro be used commercially?
The Kimi K3.1-Pro License permits commercial use, with additional terms for large Model as a Service businesses and very large commercial products. Read the license before launch.
Key takeaways
- Kimi K3.1 Pro is a 2.8T-parameter mixture-of-experts model with 104B active parameters.
- Its 96 public safetensors files total 1,561,293,160,568 bytes, about 1.45 TiB.
- The weights use MXFP4 and activations use MXFP8.
SourcesCheck the official Kimi API platform · Read the official Kimi K3.1 Pro model card · huggingface.co/moonshotai/Kimi-K3.1-Pro · huggingface.co/api/models
Last updated
Cite this pageKimi K3.1 Pro Weights: 1.45 TiB, MXFP4 and No Official VRAM Table — https://www.agentcomm.net/articles/kimi-k3-1-weights-hosting/