Only with CPU offload
Nemotron 3 Super 120B-A12B at Q4_K_M needs 70.4 GB but only 46.4 GB is addressable, so about 34% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 5.3 tokens per second — usable for batch work, painful for chat.
The open-training-data 120B. Same hybrid layout as Lightning, so 128K context costs under a gigabyte.
What hardware do I need for Nemotron 3 Super 120B-A12B? →
Fits instead: Nemotron 3.5 Lightning 30B-A3B (18.4 GB)
The VRAM budget
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 122.7 GB | 123.4 GB | — | ~1.8 | −0.1% ppl | 77.0 GB over |
| Q6_K | 94.7 GB | 95.4 GB | — | ~2.8 | −0.4% ppl | 49.0 GB over |
| Q5_K_M | 81.8 GB | 82.5 GB | — | ~3.7 | −0.8% ppl | 36.1 GB over |
| Q4_K_M | 69.7 GB | 70.4 GB | — | ~5.3 | −1.9% ppl | 24.0 GB over |
| Q3_K_M | 56.4 GB | 57.1 GB | — | ~10 | −5.4% ppl | 10.7 GB over |
| Q2_K | 48.4 GB | 49.0 GB | — | ~25 | −15% ppl | 2.6 GB over |
Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it. Only 8 of its 88 blocks keep a per-token KV cache; the rest are linear-attention, Mamba or convolution blocks with a fixed-size state.
How to run it
$ llama-server \
-hf nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16:Q4_K_M \
-c 8192 -ngl 57
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.