MiniMax M3.1 Flash 428B-A23B — 显存、速度与本地部署
作者： Jakub Rusinowski · 最后更新：
MiniMax's October refresh of M3 for coding, agents, computer use and cybersecurity: the same 428B-total, 23B-active multimodal MoE (text, image and video in, 1M-token context) with sparse attention throughout, quantisation-aware training and speculative decoding. The native checkpoint keeps its experts in NVFP4, so it downloads at 252 GB instead of M3's full-precision size. Datacentre-scale to self-host.
MiniMax M3.1 Flash 428B-A23B 在 Q4_K_M 下约需 259 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 | 141.5 GB | — | 放不下 | |
| Q3_K_M 3.41 bpw | 183.2 GB | — | 放不下 | |
| Q4_K_M 4.83 bpw | 259.2 GB | — | 放不下 | |
| Q5_K_M 5.67 bpw | 304.1 GB | — | 放不下 | |
| Q6_K 6.56 bpw | 351.8 GB | — | 放不下 | |
| Q8_0 8.50 bpw | 455.6 GB | — | 放不下 | |
| F16 16.00 bpw | 856.8 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. MiniMax M3.1 Flash 428B-A23B 显存计算器 →
运行 MiniMax M3.1 Flash 428B-A23B
目录中能运行 MiniMax M3.1 Flash 428B-A23B 的最便宜 GPU 是 Apple M3 Ultra (512 GB).
作为亚马逊联盟成员，我们从符合条件的购买中获得收入。云 GPU 链接为推荐链接——我们可能获得佣金，您无需额外付费。
如何运行 MiniMax M3.1 Flash 428B-A23B
No Ollama tag for M3.1 Flash on 2026-10-25.
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
- 428 Billion (23B active)
- 上下文窗口
- 1M
- 架构
- Mixture-of-Experts (multimodal), MiniMax Sparse Attention
- 提供商
- MiniMax
- 许可证
- MiniMax Community License
- Availability
- Weights downloadable
- Practical deployment
- Workstation (large unified memory or pro GPU)
- 规格量化
- Q4_K_M
- 系统内存
- 512 GB
- 记录更新于
- 2026-10-25
Weights are downloadable and commercial use is permitted, subject to the licence’s acceptable-use terms.
限制与注意事项
- The 252 GB figure is MiniMax's own NVFP4/MXFP8 download (summed from the Hugging Face file listing). The Q4_K_M figure on this page is sized from the parameter count, as for any GGUF build; the two are different files.
- No KV-cache figure: MiniMax Sparse Attention is not modelled here, so the per-quant table is sized at zero context.
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 MiniMax M3.1 Flash 428B-A23B 吗？
- MiniMax M3 在 AMD Ryzen AI Max+ 395 上
- MiniMax M3 在 Apple M1 Max 上
- MiniMax M3 在 Apple M1 Ultra 上
- MiniMax M3 在 Apple M2 Max 上
- MiniMax M3 在 Apple M2 Ultra 上
- MiniMax M3 在 Apple M3 Max 上
- MiniMax M3 在 Apple M4 Max 上
- MiniMax M3 在 Apple M5 Max 上
- MiniMax M3 在 Apple M5 Pro 上
- MiniMax M3 在 NVIDIA A100 80GB (PCIe) 上
- MiniMax M3 在 NVIDIA DGX Spark 上
- MiniMax M3 在 NVIDIA H100 80GB (PCIe) 上
- MiniMax M3 在 NVIDIA RTX PRO 6000 Blackwell 上
MiniMax M3 的其他尺寸
MiniMax M3.1 Flash 428B-A23B — 常见问题
How much VRAM does MiniMax M3.1 Flash 428B-A23B need?
About 259 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 MiniMax M3.1 Flash 428B-A23B run on an RTX 4090 (24 GB)?
No. MiniMax M3.1 Flash 428B-A23B needs about 259 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 MiniMax M3.1 Flash 428B-A23B locally?
No Ollama tag for M3.1 Flash on 2026-10-25. Running the published tag would send your prompts to a hosted GPU rather than your own machine.
What other sizes does MiniMax M3 come in?
MiniMax M3 428B-A23B (259 GB), MiniMax M3-VL (140 GB), MiniMax M3.1 Flash 428B-A23B (259 GB). Every size shares the family's training and licence; the larger ones score higher and need proportionally more memory.