Yes, just — 8.0 GB spare
Nemotron 3 Super 120B-A12B at Q4_K_M fits your H100 PCIe entirely on the GPU at 8K context, at an estimated 69 tokens per second. There is room for its full 256K window.
Fully on GPU
8K context
Q4_K_M · 69.7 GB
NVIDIA Open Model
Released Mar 2026
Not in the Ollama library
The open-training-data 120B. Same hybrid layout as Lightning, so 128K context costs under a gigabyte.
The VRAM budget
weights 69.7 GB
Weights 69.7 GB
KV cache @ 8K 0.06 GB
Runtime overhead 0.6 GB
Free 8.0 GB of 78.4 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 122.7 GB | 123.4 GB | — | ~3.1 | −0.1% ppl | 45.0 GB over |
| Q6_K | 94.7 GB | 95.4 GB | — | ~7.5 | −0.4% ppl | 17.0 GB over |
| Q5_K_M | 81.8 GB | 82.5 GB | — | ~22 | −0.8% ppl | 4.1 GB over |
| Q4_K_M | 69.7 GB | 70.4 GB | 256K | 69 | −1.9% ppl | Recommended |
| Q3_K_M | 56.4 GB | 57.1 GB | 256K | 85 | −5.4% ppl | Long context |
| Q2_K | 48.4 GB | 49.0 GB | 256K | 100 | −15% ppl | Long context |
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 99
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.
01Download is 69.7 GB. Keep it on an SSD — a first load off a spinning disk takes minutes.
02Close anything else holding VRAM. A browser with hardware acceleration can sit on 1–2 GB.
03Only 8.0 GB is spare, so a long system prompt can still push it over. Its real ceiling here is 256K context.