MiniMax-M2 VRAM Requirements
MiniMax-M2 needs about 218 GB of VRAM at FP8 (official) with an 8K context: 215 GB for weights, 1.9 GB for the KV cache and 0.9 GB of runtime overhead. At Q4_K_M it needs 133 GB. At Q8_0 it needs 230 GB and at full FP16 precision 431 GB. The smallest GPU that runs it fully is the Apple M3 Ultra 512GB.
218 GB VRAM
MiniMax-M2 at FP8 (official) with 8K context and the KV cache in FP16.
- Weights 215 GB
- KV cache 1.9 GB
- Runtime overhead 0.9 GB
Mixture of experts: all 230B parameters need memory, but only 10B are read per token, so it runs at the speed of a much smaller model.
Which GPUs can run it
- RTX 3060 12GBToo small
- RTX 4060 Ti 16GBToo small
- RTX 3090 24GBToo small
- Radeon RX 7900 XTX 24GBToo small
- RTX 4090 24GBToo small
- RTX 5090 32GBToo small
- RTX 6000 Ada 48GBToo small
- A100 80GB35% on GPU~5.5 tok/s
- H100 80GB SXM35% on GPU~5.5 tok/s
- Ryzen AI Max+ 395 128GBToo small
- Apple M4 Max 128GBToo small
- Apple M5 Max 128GBToo small
- RTX PRO 6000 Blackwell 96GB41% on GPU~6.1 tok/s
- NVIDIA DGX Spark 128GBToo small
- Apple M3 Ultra 512GBFits~37 tok/s
MiniMax-M2 VRAM by quantization
Total memory with an FP16 KV cache. "Smallest tier" is the smallest standard VRAM size with a 6% safety margin at 8K.
| Quantization | Bits/weight | Weights | Total at 8K | Total at 32K | Smallest tier |
|---|---|---|---|---|---|
| FP8 (official) | 8.03 | 215 GB | 218 GB | 224 GB | Multi-GPU / server |
| Q4_K_M | 4.85 | 130 GB | 133 GB | 139 GB | Multi-GPU / server |
| Q5_K_M | 5.69 | 152 GB | 155 GB | 161 GB | Multi-GPU / server |
| Q6_K | 6.59 | 176 GB | 179 GB | 185 GB | Multi-GPU / server |
| Q8_0 | 8.50 | 228 GB | 230 GB | 236 GB | Multi-GPU / server |
| FP16 / BF16 | 16.00 | 428 GB | 431 GB | 437 GB | Multi-GPU / server |
| Q3_K_M | 3.91 | 105 GB | 108 GB | 113 GB | 128 GB |
| Q2_K | 3.00 | 80.3 GB | 83.2 GB | 89.0 GB | 96 GB |
VRAM by context length
At FP8 (official). The KV cache grows linearly with context: 0.24 GB per 1,000 tokens in FP16.
| Context | KV cache (FP16) | KV cache (Q8_0) | Total, FP16 cache | Total, Q8_0 cache |
|---|---|---|---|---|
| 4K tokens | 1.0 GB | 0.5 GB | 217 GB | 216 GB |
| 8K tokens | 1.9 GB | 1.0 GB | 218 GB | 217 GB |
| 32K tokens | 7.8 GB | 4.1 GB | 224 GB | 220 GB |
| 128K tokens | 31.0 GB | 16.5 GB | 247 GB | 232 GB |
| 192K tokens | 46.5 GB | 24.7 GB | 262 GB | 241 GB |
GPUs that can run MiniMax-M2
At FP8 (official) with 8K context. "Best quant" is the highest-quality format that fits fully; "max context" is at FP8 (official).
| GPU | Usable memory | Status | Best quant | Max context | Speed |
|---|---|---|---|---|---|
| Apple M3 Ultra 512GB | 384 GB | Fits | FP8 (official) | 192K | ~37 tok/s |
| AMD Instinct MI300X 192GB | 192 GB | 83% on GPU | Q6_K | – | ~18 tok/s |
| NVIDIA H200 141GB | 141 GB | 61% on GPU | IQ4_XS | – | ~9.0 tok/s |
| RTX PRO 6000 Blackwell 96GB | 96.0 GB | 41% on GPU | Q2_K | – | ~6.1 tok/s |
| A100 80GB | 80.0 GB | 35% on GPU | – | – | ~5.5 tok/s |
| H100 80GB SXM | 80.0 GB | 35% on GPU | – | – | ~5.5 tok/s |
About MiniMax-M2
A 230B mixture-of-experts with only 10B active parameters, built for coding agents. The official weights are FP8. Every one of its 62 layers uses full attention, so the KV cache grows quickly with long agent sessions: about 248 KB per token in FP16.
| Spec | Value |
|---|---|
| Parameters | 230B total, 10B active (MoE) |
| Layers | 62 |
| Hidden size | 3072 |
| Attention heads / KV heads | 48 / 8 |
| Head dimension | 128 |
| Vocabulary | 200,064 |
| Max context (used here) | 192K tokens |
| KV cache per token (FP16) | 248 KB |
| Released | 2025-10 |
| License | MIT |
MiniMax-M2 VRAM: common questions
What do I need to run MiniMax-M2 locally?
At Q4_K_M about 130 GB plus cache, so a 192 GB or 512 GB Mac Studio, or four to six 24–48 GB GPUs. A 128 GB machine needs a Q2-level quant.
How much VRAM does MiniMax-M2 need?
About 218 GB at FP8 (official) with an 8K context, 230 GB at Q8_0 and 431 GB at FP16. Longer contexts add 0.24 GB per 1,000 tokens with an FP16 cache.
What is the cheapest GPU that can run MiniMax-M2?
The smallest listed device that fits it fully at FP8 (official) with 8K context is the Apple M3 Ultra 512GB (384 GB usable).
Does MiniMax-M2 need VRAM for all 230B parameters?
Yes. All experts must be loaded because any token can use any expert. Only 10B parameters are read per token, so it generates text about as fast as a 10B dense model.