Yes, just — 1.6 GB spare
Nemotron 3 Super 120B-A12B at Q4_K_M fits your M2 Max · 96 GB entirely in unified memory at 8K context, at an estimated 15 tokens per second. Past 214K the KV cache pushes it over — quantise the cache to q8_0, or step down a quantisation, to go longer.
Fully in unified memory
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 1.6 GB of 72.0 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 122.7 GB | 123.4 GB | — | ~8.8 | −0.1% ppl | 51.4 GB over |
| Q6_K | 94.7 GB | 95.4 GB | — | ~11 | −0.4% ppl | 23.4 GB over |
| Q5_K_M | 81.8 GB | 82.5 GB | — | ~13 | −0.8% ppl | 10.5 GB over |
| Q4_K_M | 69.7 GB | 70.4 GB | 214K | 15 | −1.9% ppl | Recommended |
| Q3_K_M | 56.4 GB | 57.1 GB | 256K | 19 | −5.4% ppl | Long context |
| Q2_K | 48.4 GB | 49.0 GB | 256K | 22 | −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
$ pip install mlx-lm $ mlx_lm.generate --model mlx-community/NVIDIA-Nemotron-3-Super-120B-A12B-BF16-4bit \ --max-tokens 512 --prompt "Hello"
Apple's own array framework. The fastest path on Apple Silicon. More on MLX.
01Download is 69.7 GB. Keep it on an SSD — a first load off a spinning disk takes minutes.
02macOS caps what the GPU may wire down at about 72.0 GB of 96 GB. Raising it with iogpu.wired_limit_mb is possible, and risky.
03Only 1.6 GB is spare, so a long system prompt can still push it over. Its real ceiling here is 214K context.