MiMo-V2.6-Pro — 显存、速度与本地部署
作者： Jakub Rusinowski · 最后更新：
A 1.02T-parameter sparse MoE (384 routed experts, 8 active) with 42B parameters active per token, a 681M-parameter vision encoder and an audio encoder, and a 5-layer multi-token-prediction drafter. 1M-token context, MIT licence. The published weights are the RL-tuned checkpoint, MiMo-V2.6-Pro-RL. Sized at Q4_K_M it needs far more memory than any single machine in our catalogue.
MiMo-V2.6-Pro 在 Q4_K_M 下约需 619 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 | 337.5 GB | — | 放不下 | |
| Q3_K_M 3.41 bpw | 437.4 GB | — | 放不下 | |
| Q4_K_M 4.83 bpw | 619.2 GB | — | 放不下 | |
| Q5_K_M 5.67 bpw | 726.7 GB | — | 放不下 | |
| Q6_K 6.56 bpw | 840.6 GB | — | 放不下 | |
| Q8_0 8.50 bpw | 1089 GB | — | 放不下 | |
| F16 16.00 bpw | 2049.2 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. MiMo-V2.6-Pro 显存计算器 →
运行 MiMo-V2.6-Pro
MiMo-V2.6-Pro is not practical on a single consumer GPU: it needs about 619.2 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 MiMo-V2.6-Pro
No Ollama library entry. Xiaomi's documented serving routes are SGLang and vLLM.
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
- 1.02 Trillion (42B active)
- 上下文窗口
- 1,048,576
- 架构
- Sparse MoE (384 experts, 8 active), hybrid sliding-window attention; vision + audio encoders
- 提供商
- Xiaomi
- 许可证
- MIT
- Availability
- Weights downloadable
- Practical deployment
- Private cloud / datacenter cluster
- 规格量化
- Q4_K_M
- 系统内存
- 1024 GB
- 记录更新于
- 2026-10-13
Commercial use permitted. No usage restrictions beyond attribution.
另请参阅 MiMo-V2.5-Pro 1T
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。
Independent evaluation — not vendor-claimed and not measured by us
| 基准 | 分数 | 来源 |
|---|---|---|
| AA Intelligence Index v4.3.2 | 46 | independent · Artificial Analysis · snapshot 2026-10-13 |
MiMo-V2.6 的其他尺寸
MiMo-V2.6-Pro — 常见问题
How much VRAM does MiMo-V2.6-Pro need?
About 619 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 MiMo-V2.6-Pro run on an RTX 4090 (24 GB)?
No. MiMo-V2.6-Pro needs about 619 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 MiMo-V2.6-Pro locally?
No Ollama library entry. Xiaomi's documented serving routes are SGLang and vLLM. Running the published tag would send your prompts to a hosted GPU rather than your own machine.
What other sizes does MiMo-V2.6 come in?
MiMo-V2.6-Pro (619 GB), MiMo-V2.6-Flash (188 GB), MiMo-V2.6 Distill 9B (6 GB). Every size shares the family's training and licence; the larger ones score higher and need proportionally more memory.