MiniMax M3
MiniMax M3 needs roughly 293.1 GB VRAM at Q4_K_M quantization (959.9 GB at FP16). 6 GPUs we track can run it fully in VRAM at 8k context.
6 GPUs run this natively · 1 with CPU offload
MiniMax M3 is a Mixture of Experts (MoE) large language model with 428B total parameters but only 23B active per token developed by MiniMax. Released in June 2026, it supports text and vision and video inputs with a 1M context window, released under the MiniMax Community license, allowing commercial use. Native multimodal MoE using MiniMax Sparse Attention to make 1M-token inference practical, with enabled, adaptive, and disabled reasoning modes. MiniMax reports SWE-bench Verified 80.5, SWE-bench Pro 59.0, and Terminal-Bench 2.1 66.0. NVIDIA provides an NVFP4 conversion, but local deployment remains a multi-GPU server workload.
To run MiniMax M3 locally, you need approximately 293.1 GB of VRAM at Q4_K_M quantization with 8k context. 6 of the GPUs we track can run it fully in VRAM, with a further 1 able to offload to system RAM. At Q4_K_M it requires 293.1 GB — more than any single consumer GPU. A multi-GPU server or 80 GB datacenter GPU is required. At Q8_K_M (510.7 GB), you get near-FP16 quality while still fitting on large server or multi-GPU setups. FP16 requires 959.8 GB, limiting it to datacenter-class hardware with 80 GB+ VRAM. As a MoE model, inference speed depends on active parameters (23B) rather than total size, so once it fits in VRAM it runs noticeably faster than dense models of the same parameter count.
MiniMax M3 is designed for reasoning and problem-solving. Thanks to its MoE architecture with only 23B active parameters, it delivers strong quality while keeping inference latency manageable for local use.
VRAM at each quantization
MiniMax M3 natively supports a longer context window, but the table below is capped at 8k for comparability — KV cache grows linearly with context length.
| Quant | Weights | KV cache | Total |
|---|---|---|---|
| FP32 | 1712.0 GB | 1.01 GB | 1918.6 GB |
| BF16 | 856.0 GB | 1.01 GB | 959.9 GB |
| FP16 | 856.0 GB | 1.01 GB | 959.9 GB |
| Q8_0 | 455.0 GB | 1.01 GB | 510.7 GB |
| Q6_K | 351.4 GB | 1.01 GB | 394.7 GB |
| Q5_K_M | 304.7 GB | 1.01 GB | 342.4 GB |
| Q4_K_M | 260.6 GB | 1.01 GB | 293.1 GB |
| Q3_K_M | 205.9 GB | 1.01 GB | 231.7 GB |
| Q2_K | 163.1 GB | 1.01 GB | 183.8 GB |
| NVFP4reccuda | 214.0 GB | 1.01 GB | 240.8 GB |
KV cache figures assume 8k context at FP16. NVFP4 quantization requires a CUDA-capable GPU. Enable TurboQuant in the calculator to see reduced KV cache estimates.
Benchmarks
GPUs that run MiniMax M3 natively (6)
- Apple M4 Ultra (384GB)Q5_K_M · 15.7 t/s
- Apple M4 Ultra (192GB)Q2_K · 28.9 t/s
- Apple M3 Ultra (512GB)Q6_K · 10.2 t/s
- Apple M3 Ultra (256GB)Q3_K_M · 17.3 t/s
- Apple M2 Ultra (384GB)Q5_K_M · 11.5 t/s
- Apple M2 Ultra (192GB)Q2_K · 21.2 t/s
Plus 1 GPUs that run it with CPU offload (slower)
- AMD Instinct MI300XQ2_K · 114 t/s
Notes
Native multimodal MoE using MiniMax Sparse Attention to make 1M-token inference practical, with enabled, adaptive, and disabled reasoning modes. MiniMax reports SWE-bench Verified 80.5, SWE-bench Pro 59.0, and Terminal-Bench 2.1 66.0. NVIDIA provides an NVFP4 conversion, but local deployment remains a multi-GPU server workload.
How to run MiniMax M3 locally
NVFP4 needs 240.8 GB — needs multiple datacenter-class GPUs (80 GB+ each).
vLLM
vllm serve MiniMaxAI/MiniMax-M3 --trust-remote-codeLM Studio: This is a multi-GPU server model. NVIDIA's NVFP4 conversion is the practical low-bit path; generic desktop GGUF support should not be assumed to reproduce MiniMax Sparse Attention efficiently at 1M context.
Why this quantization? NVFP4 cuts the 428B weight pool to roughly 214 GB and has a published NVIDIA deployment path. M3's sparse attention reduces long-context compute and KV work, but it does not make the full expert weights fit on consumer hardware.
Who is MiniMax M3 for?
Teams building multimodal coding, computer-use, or office agents on multi-GPU servers and needing a genuinely long 1M-token window.
Best for
- Long-horizon software engineering
- Image/video-grounded agents
- Large-repository work
- Adaptive reasoning workloads
Not ideal for
- Single consumer GPUs
- Small low-latency assistants
- Deployments that cannot run custom sparse-attention kernels
Frequently asked questions
- What are the VRAM requirements for MiniMax M3?
- MiniMax M3 requires approximately 293.1 GB of VRAM at Q4_K_M quantization, 510.7 GB at Q8, and 959.8 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 M3 have?
- MiniMax M3 has 428 billion total parameters, but only 23 billion are active per token thanks to its Mixture of Experts (MoE) architecture. This makes inference significantly faster than the total parameter count suggests.
- Can MiniMax M3 run on a 16 GB GPU?
- No. At Q4_K_M, MiniMax M3 needs 293.1 GB of VRAM — more than 16 GB. You will need a multi-GPU server.
- Can MiniMax M3 run on a 24 GB GPU?
- No. Even at Q4_K_M, MiniMax M3 needs 293.1 GB. Consider a multi-GPU server with 80 GB+ total VRAM.
- What is the smallest quantization for MiniMax M3 that fits in 24 GB of VRAM?
- MiniMax M3 cannot fit in 24 GB of VRAM at any standard quantization level. The minimum needed is 183.8 GB at Q2_K.
- What GPU do I need to run MiniMax M3 locally?
- You need a multi-GPU server. At Q4_K_M, MiniMax M3 needs 293.1 GB VRAM, more than any single consumer GPU. Consider 2–4× H100 or A100 GPUs.