TechCompare LogoTechCompare

How much VRAM does Nemotron 3.5 Lightning 30B need at Q4_K_M? NVIDIA's single-card 1M model

Nemotron 3.5 Lightning runs the full 1M context on a single 24 GB card with active-only loading (~20 GB), which no comparable 1M-window model manages. Resident mode is 36 GB, so a 48 GB pro card runs it with all experts warm.

Nemotron 3.5 Lightning 30B at Q4_K_M needs just 36 GB of VRAM with all experts resident at the full 1M context, or 20 GB with active-expert offload. NVIDIA's hybrid Mamba-2 plus transformer design with only 3B active parameters per token is what makes a 1M-context model land inside a single 24 GB budget - the reason it's the hottest local agent model on HuggingFace right now.

By TechCompare · Updated

Total VRAM required
37.4 GB
Nemotron 3.5 Lightning 30B (MoE) at Q4_K_M
Weights
16.8 GB
30B params
KV cache
17.2 GB
1024K tokens, FP16 KV

Calculator

Estimated VRAM required

37.4 GB

30B params at Q4_K_M, 1,048,576 token context, batch 1, inference.

Weights
16.8 GB
KV cache
17.2 GB
Overhead
3.4 GB
Doesn't fit on a 32 GB consumer GPU at Q4_K_M. Try (29.5 GB) for the smallest quant that fits a single RTX 5090.

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.

Hardware that fits

A100 40GB
Datacenter
40 GB
93% used
Apple M3 Max 64GB
Unified
48 GB
78% used
RTX 6000 Ada
Pro
48 GB
78% used

Just barely too small

RTX 5090
Consumer
32 GB
short by 5.4 GB

How this is calculated

Only 8 attention layers carry a key-value cache; the rest of the transformer blocks are Mamba-2 state-space layers with a fixed-size recurrent state. Weights at Q4_K_M are 16.8 GB, the KV cache with 4 KV heads at head_dim 128 across 8 attention layers is 16.4 GB at 1M, and overhead adds 3.3 GB. Active-only keeps just the 3B routed experts resident (1.7 GB) so the total drops to ~20 GB, fitting a 24 GB consumer card with room for batch.

Verdict

The math is what sells NVIDIA's hybrid bet: 16.8 GB of Q4 weights plus a 16.4 GB KV cache (only 8 attention layers contribute, and the Mamba-2 blocks hold a fixed state instead of growing a cache) plus 3.3 GB of overhead totals 36 GB resident. Active-only drops weights to 1.7 GB of routed experts and the total to ~20 GB. Compared to Nemotron 3 Ultra 550B - the other hybrid on this list - Lightning trades capacity for a footprint a single RTX 4090 or 5090 handles outright.

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

How can a 1M-context model fit in 20 GB?
Most of the transformer blocks are Mamba-2 state-space layers, which keep a fixed-size recurrent state instead of growing a key-value cache. Only 8 attention layers accumulate KV. Combined with just 3B active parameters per token, active-only loading lands at 1.7 GB of weights plus 16.4 GB of cache plus a couple GB of overhead.
Nemotron 3.5 Lightning or gpt-oss 20B for a local agent?
Lightning for long-context agentic work: 1M native window at a similar footprint. gpt-oss 20B for maximum compatibility and the cleanest tooling, but its window is 128K. On a 24 GB card both run; Lightning uses the memory budget mostly for cache, gpt-oss for weights.