MiniMax M1 456B
MiniMax M1 456B needs roughly 290.5 GB VRAM at Q4_K_M quantization (1024.5 GB at FP16). 7 GPUs we track can run it fully in VRAM at 8k context.
7 GPUs run this natively · 0 with CPU offload
MiniMax M1 456B is a Mixture of Experts (MoE) model with 456B total parameters but only 46B active per token developed by MiniMax. June 2025 456B MoE with 46B active parameters and 1M context. Hybrid lightning attention (linear+softmax).
To run MiniMax M1 456B locally: Q2_K ~150-180GB — multi-GPU server only. The 40k/80k thinking-budget variants are more practical. As a MoE model, inference speed depends on active parameters (46B) rather than total size.
MMLU-Pro 80.6%, GPQA 69.2%, Math 96.0% — frontier reasoning with massive context.
VRAM at each quantization
Assumes 8k context. KV cache grows linearly with context length.
| Quant | Weights | KV cache | Total |
|---|---|---|---|
| FP32 | 1824.0 GB | 2.68 GB | 2045.9 GB |
| BF16 | 912.0 GB | 2.68 GB | 1024.5 GB |
| FP16 | 912.0 GB | 2.68 GB | 1024.5 GB |
| Q8_0 | 456.0 GB | 2.68 GB | 513.7 GB |
| Q6_K | 373.9 GB | 2.68 GB | 421.8 GB |
| Q5_K_M | 293.7 GB | 2.68 GB | 331.9 GB |
| Q4_K_M | 256.7 GB | 2.68 GB | 290.5 GB |
| Q3_K_M | 196.1 GB | 2.68 GB | 222.6 GB |
| Q2_Krec | 150.0 GB | 2.68 GB | 171.0 GB |
| NVFP4cuda | 228.0 GB | 2.68 GB | 258.4 GB |
KV cache shown at 8k context (FP16). NVFP4 requires a CUDA GPU. Enable TurboQuant in the calculator to see reduced KV cache estimates.
Benchmarks
GPUs that run MiniMax M1 456B natively (7)
- AMD Instinct MI300XQ2_K · 385.2 t/s
- Apple M4 Ultra (384GB)Q5_K_M · 40.5 t/s
- Apple M4 Ultra (192GB)Q2_K · 79.4 t/s
- Apple M3 Ultra (512GB)Q6_K · 23.9 t/s
- Apple M3 Ultra (256GB)Q3_K_M · 45.5 t/s
- Apple M2 Ultra (384GB)Q5_K_M · 29.7 t/s
- Apple M2 Ultra (192GB)Q2_K · 58.1 t/s
Notes
Hybrid lightning attention (linear+softmax) reasoning model with 1M context. 40k and 80k thinking-budget variants available.
Frequently asked questions
- What are the VRAM requirements for MiniMax M1 456B?
- MiniMax M1 456B requires approximately 290.5 GB of VRAM at Q4_K_M quantization, 513.7 GB at Q8, and 1024.4 GB at FP16. These numbers assume 8k context window; VRAM scales linearly with context length due to the KV cache.
- How many parameters does MiniMax M1 456B have?
- MiniMax M1 456B has 456 billion total parameters, but only 46 billion are active per token thanks to its Mixture of Experts (MoE) architecture. This makes inference significantly faster than the total parameter count suggests.
- Is MiniMax M1 456B good at reasoning and math?
- Yes. With a MATH score of 96 and MMLU-Pro of 81.1, MiniMax M1 456B handles complex multi-step reasoning, analytical tasks, and problem-solving well.
- Can MiniMax M1 456B run on a 16 GB GPU?
- No. At Q4_K_M, MiniMax M1 456B needs 290.5 GB of VRAM — more than 16 GB. You will need a multi-GPU server.
- Can MiniMax M1 456B run on a 24 GB GPU?
- No. Even at Q4_K_M, MiniMax M1 456B needs 290.5 GB. Consider a multi-GPU server with 80 GB+ total VRAM.
- What is the smallest quantization for MiniMax M1 456B that fits in 24 GB of VRAM?
- MiniMax M1 456B cannot fit in 24 GB of VRAM at any standard quantization level. The minimum needed is 171.0 GB at Q2_K.
- What GPU do I need to run MiniMax M1 456B locally?
- You need a multi-GPU server. At Q4_K_M, MiniMax M1 456B needs 290.5 GB VRAM, more than any single consumer GPU. Consider 2–4× H100 or A100 GPUs.