Can I Run Nemotron 3 Super on 96 GB system RAM?

作者： Jakub Rusinowski · 最后更新： 2026年8月15日

Yes, but it is tight

Yes, but it is tight — Nemotron 3 Super 120B-A12B at Q4_K_M needs about 76.1 GB of the 76.8 GB usable on 96 GB system RAM, leaving only ~0.7 GB before the runtime starts swapping. Expect ~7.6 tok/s (estimated), with room for about 8,192 tokens of context.

Confidence: low · Recommended quantization: Q4_K_M · Estimated speed: ~7.6 tok/s

Buy This HardwareRyzen AI Max+ 395 Laptop (Strix Halo, up to 128GB) — 96 GB VRAM · 120 W board power Deploy in the Cloud NowNVIDIA A100 80GB on RunPod — from $1.39/hr · rate checked 2026-07

or compare on Vast.ai from $0.77/hr (typical low · varies)

As an Amazon Associate we earn from qualifying purchases. Cloud GPU links are referral links — we may earn a commission at no extra cost to you.

96 GB system RAM — what it gives a model

Usable memory for models	76.8 GB
Memory bandwidth	90 GB/s

Nemotron 3 Super on 96 GB system RAM: memory by quantization

Quant	Memory needed	Fits 76.8 GB?	Max context	Est. speed	Download
F16	243.7 GB	✗ No	—	—	240 GB
Q8_0	131.2 GB	✗ No	—	—	127.5 GB
Q6_K	102.1 GB	✗ No	—	—	98.4 GB
Q5_K_M	88.7 GB	✗ No	—	—	85.1 GB
Q4_K_M	76.1 GB	✓ Yes	8K	~7.6 tok/s	72.5 GB
Q3_K_M	54.8 GB	✓ Yes	64K	~10 tok/s	51.2 GB
Q2_K	43.1 GB	✓ Yes	64K	~12.2 tok/s	39.4 GB

What to watch out for

Only ~0.7 GB of headroom at Q4_K_M: a longer context or a second application can push this into swapping.
This model does not publish its full attention configuration, so the KV-cache share of these figures is inferred from its parameter count rather than computed exactly.
These figures assume CPU-only inference. Any discrete GPU, even an 8 GB one, will be several times faster for models that fit in its VRAM.

Recommended setup

llama.cpp (CPU build) or Ollama — both run without a GPU

How these numbers are calculated

Memory need = quantized weights + KV cache at 8,192 tokens (f16) + 0.8 GB runtime overhead.
76.8 GB of the 96 GB is treated as usable for model weights (80% — the rest is the OS and running applications).
DDR5-5600 dual channel at 89.6 GB/s peak. CPU decode is assumed to sustain 35% of that peak, because CPU inference is not purely bandwidth-bound — it also spends real time in compute and thread synchronisation. This figure is an assumption, not a fitted constant: no CPU measurement is in the calibration set.
CPU-only inference: no GPU is assumed. A GPU of any size will beat these figures substantially.
Throughput is a memory-bandwidth roofline estimate, not a measurement. It is labelled "estimated" everywhere it appears.
KV cache is inferred from the parameter count because this model does not publish its full attention configuration.

FAQ