How much VRAM does MiniMax-M3 428B (MoE) need at Q4_K_M? MiniMax 1M block-sparse

MiniMax-M3 428B at Q4_K_M is a 405 GB resident deployment for 1M context. Active-only only gets you to 156 GB because the full-resolution key-value cache at 1M context is the binding constraint.

MiniMax-M3 428B at Q4_K_M with native 1M context needs about 405 GB of VRAM with all experts resident. MiniMax's third generation activates 23B parameters per token from a 428B total pool. The architecture uses block-sparse MSA for attention compute, but stores the key-value cache at full resolution. Active-expert offload drops the resident footprint to roughly 156 GB.

By TechCompare · Updated

Total VRAM required
405 GB
MiniMax-M3 428B (MoE) at Q4_K_M
Weights
240 GB
428B params
KV cache
129 GB
1024K tokens, FP16 KV

Calculator

Estimated VRAM required

405 GB

428B params at Q4_K_M, 1,048,576 token context, batch 1, inference.

Weights
240 GB
KV cache
129 GB
Overhead
36.9 GB

Estimate accuracy: Weights within ~2%. KV cache within ~5% for standard GQA models, ~10% for MLA (DeepSeek). Real VRAM may vary with framework (vLLM vs llama.cpp vs Transformers), Flash Attention, and driver overhead.

Hardware that fits

No single GPU in our catalog has enough memory. Multi-GPU or CPU offload required.

How this is calculated

The 428B parameter pool requires 240 GB of weights at Q4_K_M. The 60-layer architecture consists of 3 dense layers plus 57 MoE layers, using group-query attention with 4 KV heads at head_dim 128. Block-sparse MSA is an attention compute optimization that attends to the top-16 blocks of 128 tokens, not a storage compression, so the key-value cache is stored full-resolution. At the full 1M context the FP16 key-value cache uses a significant 129 GB. Activation and software overhead adds about 37 GB. In active-only mode the resident weights shrink to 13 GB, but the 129 GB cache stays, which drops total usage to 156 GB.

Verdict

Resident mode needs six 80GB datacenter cards or three 141GB H200 cards. Active-only at 156 GB fits on dual 80GB cards, but the 1M-context cache dominates. Block-sparse MSA speeds up attention compute without shrinking the cache, so capping context is the fastest way to reduce footprint. The MiniMax API is the economical option for most workloads.

More MiniMax scenarios

DeepSeek V4 Pro 1.6T (MoE) at Q4_K_M
1600B - Q4_K_M - 1024K ctx
View details ➜
Llama 4 Scout (17B/109B) at Q4_K_M
109B - Q4_K_M - 10240K ctx
View details ➜
gpt-oss 20B (MoE) at Q4_K_M
20B - Q4_K_M - 128K ctx
View details ➜

Frequently asked questions

Does block-sparse MSA reduce the key-value cache?
No. Block-sparse MSA optimizes attention compute by attending only to the top-16 blocks of 128 tokens, which speeds up the attention pass. The key-value cache is still stored at full sequence resolution so future queries can attend to any past token. The cache size is the same as standard group-query attention.
How does MiniMax-M3 compare to M2.7?
M3 has a 428B total pool with 23B active across 60 layers versus M2.7's 230B pool across 64 layers. M3 upgrades to a native 1M context window and block-sparse MSA compute. The M2.7 cache at 256K context was smaller. The M3 cache at 1M context is about 129 GB due to the longer window, not the architecture.