TechCompare LogoTechCompare

How much VRAM does Nemotron 3 Super 120B need at Q4_K_M? NVIDIA's serving-tuned MoE

Pick Nemotron 3 Super 120B in resident mode when you have datacenter hardware and predictable sustained throughput matters - it's the MoE tuned for batched serving rather than peak single-stream speed. For desktop experimentation, active-only works at the cost of variable per-token latency.

Nemotron 3 Super 120B at Q4_K_M with native 1M context needs about 376 GB of VRAM with all experts resident. It's an MoE (12B active per token / 120B total) that NVIDIA has tuned aggressively for H100/H200 batched-serving throughput - the architecture is 64 layers, hidden 6144, 8 KV heads, head_dim 128. Active-only loading drops the resident footprint to ~310 GB, but Nemotron's whole point is sustained server throughput, which is a resident-mode concern.

By TechCompare · Updated

Total VRAM required
376 GB
Nemotron 3 Super 120B (MoE) at Q4_K_M
Weights
67.2 GB
120B params
KV cache
275 GB
1024K tokens, FP16 KV

Calculator

Estimated VRAM required

376 GB

120B params at Q4_K_M, 1,048,576 token context, batch 1, inference.

Weights
67.2 GB
KV cache
275 GB
Overhead
34.2 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.

KV cache exceeds model weights: Consider lowering the context length to save on VRAM. Contexts between 8K and 64K are generally more typical for local setups.

Custom architecture - SWA not applied. If you're modeling Gemma 3/4 or Mistral Nemo, pick the preset for accurate KV cache.

Hardware that fits

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

How this is calculated

67 GB of weights at Q4_K_M (full 120B pool), 275 GB KV cache at 1M context with standard 8-KV-head GQA, and ~34 GB overhead. The 376 GB resident total fits multi-GPU hardware (such as four 141 GB or 192 GB datacenter cards with tensor parallelism). Active-only is ~310 GB, but at the usual cold-expert PCIe penalty.

Verdict

The resident math is 67 GB of Q4 weights, 275 GB of FP16 KV cache (8 KV heads at head_dim 128 across 64 layers at the full 1M context), and 34 GB of overhead totaling 376 GB. Active-only loads only the active 12B of weights, dropping that figure to 6.7 GB while the 275 GB cache stays put, for a 310 GB total. NVIDIA's H100/H200 kernel tuning is the real story: the routing, fused attention, and tensor-pipeline splits are tuned to push batched serving throughput rather than per-stream latency. The 275 GB cache at 1M context is the binding constraint for desktop setups, so capping context to 64K drops the resident footprint to about 90 GB and makes single-card operation realistic.

More Nemotron scenarios

DeepSeek V4 Pro 1.6T (MoE) at Q4_K_M
DeepSeek V4 Pro 1.6T at Q4_K_M with the full 1M-token context needs about 1012 GB of VRAM with every expert resident - this is the real shape of the model and the number to plan a deployment against.
View details ➜
Llama 4 Scout (17B/109B) at Q4_K_M
Llama 4 Scout at Q4_K_M with its native 10M context needs about 2231 GB of VRAM with all 109B params resident - that's the number you size hardware against.
View details ➜
gpt-oss 20B (MoE) at Q4_K_M
gpt-oss 20B at Q4_K_M with native 128K context needs about 19.4 GB of VRAM with all experts resident, dropping to roughly 9.3 GB with active-only weight loading.
View details ➜

Frequently asked questions

Why is Nemotron 3 Super tuned by NVIDIA for serving?
NVIDIA optimizes the kernel-level expert routing, fused attention, and tensor/pipeline parallel splits specifically for H100/H200. The result is markedly higher batched-throughput than a generic MoE of the same active-param count - particularly important for high-QPS endpoints rather than single-stream chat.
Can Nemotron 3 Super fit on a 4090/5090?
No. At native 1M context, the 275 GB KV cache alone vastly exceeds consumer GPU limits. To run the model at full 1M context, pooled datacenter cards are required. Even with active-only offload, the model requires ~310 GB of memory.
Why is Nemotron 3 Super tuned by NVIDIA for serving?
NVIDIA's inference serving optimization team tuned the architecture and quantization recipe for throughput on their H100 and H200 platforms. The MoE design (large total parameter pool, smaller active subset per token) is specifically a serving play: more total capacity for harder reasoning without proportionally more compute per token. Active-only loading keeps the per-token cost down.
What does a serving-tuned model mean for on-prem deployments?
It means the reference numbers assume batch inference on datacenter cards, not single-user interactive chat. A serving-tuned MoE sustains high tokens-per-second per dollar at large batch sizes, so the economics favor shared hosting. A solo developer running one session at a time gets less of the payoff and feels more of the cold-expert latency.