Kimi K2.6 — 显存、速度与本地部署
作者： Jakub Rusinowski · 最后更新：
April 2026 successor to K2.5, same 1T-parameter MoE shape: 384 experts with 8 routed plus 1 shared, 61 layers, MLA attention, 256K context, Modified MIT. Adds video alongside image input through MoonViT. 80.2% on SWE-bench Verified and 66.7% on Terminal-Bench 2.0 — the strongest open-weight coder at release. Like K2.5 it needs ~605 GB resident at Q4_K_M: the 32B figure is active parameters per token, not what fits on your card.
Kimi K2.6 在 Q4_K_M 下约需 605 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。
按量化级别的显存与速度
计算基准：NVIDIA RTX 4090 (24 GB)。仅含权重与开销：该模型架构未公开，因此未计入 KV 缓存。
| 量化 | 显存 | 显存 | 速度（估算） | 适配 |
|---|---|---|---|---|
| Q2_K 2.63 bpw | 329.6 GB | — | 放不下 | |
| Q3_K_M 3.41 bpw | 427.1 GB | — | 放不下 | |
| Q4_K_M 4.83 bpw | 604.6 GB | — | 放不下 | |
| Q5_K_M 5.67 bpw | 709.6 GB | — | 放不下 | |
| Q6_K 6.56 bpw | 820.8 GB | — | 放不下 | |
| Q8_0 8.50 bpw | 1063.3 GB | — | 放不下 | |
| F16 16.00 bpw | 2000.8 GB | — | 放不下 |
黑色标记 = NVIDIA RTX 4090 (24 GB) 上的可用显存。 估算来自内存带宽屋顶线模型，详见 方法说明页. Kimi K2.6 显存计算器 →
运行 Kimi K2.6
Kimi K2.6 is not practical on a single consumer GPU: it needs about 604.6 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 K2.6
安装 Ollama，然后运行：
ollama run hf.co/moonshotai/Kimi-K2.6-Instruct-Q4_K_M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
- 1 Trillion (32B active)
- 上下文窗口
- 256,000
- 架构
- MoE (384 experts, 8+1 shared, MLA) + MoonViT vision
- 提供商
- Moonshot AI
- 许可证
- Modified MIT
- Practical deployment
- Private cloud / datacenter cluster
- 规格量化
- Q4_K_M
- 系统内存
- 768 GB
- 记录更新于
- 2026-09-06
Commercial use permitted. No usage restrictions beyond attribution.
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。
| 基准 | 分数 | 来源 |
|---|---|---|
| SWE-bench Verified | 80.2 / 100 % | 已发布 · https://benchlm.ai/models/kimi-2-6 |
| Terminal-Bench 2.0 | 66.7 / 100 % | 已发布 · https://benchlm.ai/models/kimi-2-6 |
Kimi K2.5 / K2.6 / K2.7 的其他尺寸
Kimi K2.6 — 常见问题
How much VRAM does Kimi K2.6 need?
About 605 GB at Q4_K_M — quantized weights plus framework overhead, before any KV cache. The cache grows with context length and is added on top; the table above folds it in. Apple Silicon counts unified memory toward the same figure.
Does Kimi K2.6 run on an RTX 4090 (24 GB)?
No. Kimi K2.6 needs about 605 GB at Q4_K_M, more than a single RTX 4090's 24 GB. It needs a larger card, several GPUs, or Apple Silicon with enough unified memory — or it runs with part of the weights offloaded to system RAM, which is much slower.
How do I run Kimi K2.6 locally?
Install Ollama and run `ollama run hf.co/moonshotai/Kimi-K2.6-Instruct-Q4_K_M`. That pulls the weights and starts a local OpenAI-compatible endpoint; after the download nothing leaves the machine.
What other sizes does Kimi K2.5 / K2.6 / K2.7 come in?
Kimi K2.5 (605 GB), Kimi K2.6 (605 GB), Kimi K2.5 1T (32B Active) (605 GB), Kimi K2.7 Code (605 GB). Every size shares the family's training and licence; the larger ones score higher and need proportionally more memory.