Kimi K3.1 Pro — 显存、速度与本地部署
作者： Jakub Rusinowski · 最后更新：
Moonshot AI's open-weight flagship, post-trained from Kimi K3: a 2.8T-parameter mixture-of-experts with 104B active per token, a 1M-token context and image input. Its 93 layers mix linear (KDA) and gated sparse attention. The weights ship in MXFP4 from quantisation-aware training, 1.56 TB on disk, which puts it on a multi-node server. Custom licence: large hosted-API operators need a separate agreement, and very large products must display the model name.
Kimi K3.1 Pro 在 MXFP4 下约需 1562 GB 显存——量化权重加框架开销，不含 KV 缓存。在 Apple Silicon 上，这部分来自统一内存。
运行 Kimi K3.1 Pro
Kimi K3.1 Pro is not practical on a single consumer GPU: it needs about 1562.1 GB at Q4_K_M, more than any single device in our catalogue holds with headroom. How multi-GPU inference works, or rent a multi-GPU server by the hour:
如何运行 Kimi K3.1 Pro
No local Ollama tag. Ollama lists kimi-k3.1-pro only as a cloud-hosted ':cloud' tag, which runs on Ollama's servers, not your machine.
规格
Verified — Checked against the primary source — the model card or the vendor spec page — and corroborated by a second independent source.
- 参数量
- 2.8 Trillion (104B active)
- 上下文窗口
- 1,048,576
- 架构
- Mixture-of-Experts (896 experts, 16 active + 2 shared), 69 KDA + 24 Gated DSA layers, MoonViT-V2 vision encoder
- 提供商
- Moonshot AI
- 许可证
- Kimi K3.1-Pro License
- Availability
- Weights downloadable
- Practical deployment
- Private cloud / datacenter cluster
- 规格量化
- MXFP4
- 系统内存
- 2048 GB
- 记录更新于
- 2026-11-11
Weights are downloadable and commercial use is permitted, subject to the licence’s acceptable-use terms.
Model-as-a-service operators whose group revenue exceeds $20M over any 12 months need a separate agreement with Moonshot AI; products above 100M monthly active users or $20M monthly revenue must display "Kimi K3.1-Pro". Internal use is exempt.
限制与注意事项
- Speed and KV cache are not modelled: the gated sparse-attention layers carry an index cache our model cannot express.
- Moonshot did not publish weights for the Highspeed variant.
另请参阅 Kimi K3 2.8T
质量与使用场景
评分由模型作者或独立评测方发布——衡量质量而非吞吐量，并非我们实测。
我的 GPU 能运行 Kimi K3.1 Pro 吗？
Kimi K3 的其他尺寸
Kimi K3.1 Pro — 常见问题
How much VRAM does Kimi K3.1 Pro need?
About 1562 GB at MXFP4 — quantized weights plus framework overhead, before any KV cache. We publish no context-length figure or speed for it: it interleaves linear (KDA) layers with gated sparse-attention layers that carry an index cache, which our memory and speed model cannot yet express, so we state the published weights only. Apple Silicon counts unified memory toward the same figure.
Does Kimi K3.1 Pro run on an RTX 4090 (24 GB)?
Its weights and overhead come to about 1562 GB at MXFP4, but we do not publish a fit verdict or a speed for it: it interleaves linear (KDA) layers with gated sparse-attention layers that carry an index cache, which our memory and speed model cannot yet express, so we state the published weights only.
How do I run Kimi K3.1 Pro locally?
No local Ollama tag. Ollama lists kimi-k3.1-pro only as a cloud-hosted ':cloud' tag, which runs on Ollama's servers, not your machine. Running the published tag would send your prompts to a hosted GPU rather than your own machine.
What other sizes does Kimi K3 come in?
Kimi K3 2.8T (1691 GB), Kimi K3.1 Pro (1562 GB). Every size shares the family's training and licence; the larger ones score higher and need proportionally more memory.