How much VRAM does Nemotron 3 Ultra 550B A55B (MoE) need at Q4_K_M? NVIDIA 1M Mamba-2

Nemotron 3 Ultra 550B at Q4_K_M is a 353 GB resident deployment for 1M context, or about 48 GB with active-expert offload. The Mamba-2 layers are why the cache is tiny enough for active-only to almost fit on a single card.

Nemotron 3 Ultra 550B at Q4_K_M with native 1M context needs about 353 GB of VRAM with all experts resident. NVIDIA's hybrid Mamba-2 plus MoE flagship activates 55B parameters per token from a 550B total pool. Only 12 of 108 transformer blocks carry attention and the other 48 are Mamba-2 state-space layers with zero key-value cache. Active-expert offload drops the resident footprint to roughly 48 GB.

By TechCompare · Updated

Total VRAM required
353 GB
Nemotron 3 Ultra 550B A55B (MoE) at Q4_K_M
Weights
308 GB
550B params
KV cache
12.9 GB
1024K tokens, FP16 KV

Calculator

Estimated VRAM required

353 GB

550B params at Q4_K_M, 1,048,576 token context, batch 1, inference.

Weights
308 GB
KV cache
12.9 GB
Overhead
32.1 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 550B parameter pool requires 308 GB of weights at Q4_K_M. The architecture is a hybrid: 12 attention layers, 48 Mamba-2 state-space layers, and 48 MoE expert blocks. Only the 12 attention layers contribute to the key-value cache, which is just 13 GB at the full 1M context with 2 KV heads at head_dim 128. The Mamba-2 layers maintain a fixed recurrent state instead of growing a cache. Activation and software overhead adds about 32 GB. In active-only mode the resident weights shrink to 31 GB, which drops total usage to 48 GB. The tiny cache means active-only is remarkably efficient for a 550B model.

Verdict

Resident mode needs five 80GB datacenter cards or three 141GB H200 cards. Active-only at 48 GB fits on a single 80GB card with headroom for batch. The 48 Mamba-2 layers replace what would otherwise be a 100 GB-plus key-value cache at 1M context, which is why this is the most single-card-friendly 500B-class MoE on the list.

More Nemotron 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

Why is Nemotron 3 Ultra's key-value cache so small?
48 of 108 transformer blocks are Mamba-2 state-space layers. Mamba-2 maintains a fixed-size recurrent state instead of growing a key-value cache with context length. Only the 12 attention layers contribute to the cache. At 1M context the cache is about 13 GB, roughly an order of magnitude smaller than a pure-attention model of the same layer count.
How does Nemotron 3 Ultra compare to Nemotron 3 Super?
Super is 120B total with 12B active across 64 attention layers at 8 KV heads, which gives a 275 GB cache at 1M context. Ultra is 550B total with 55B active and only 12 attention layers plus 48 Mamba-2 layers, which gives a 13 GB cache at 1M context. Ultra has 4.6x the parameter count but a smaller key-value cache, and the active-only footprint is 48 GB versus Super's 310 GB.