Only with CPU offload
Nemotron 3.5 Lightning 30B-A3B at Q4_K_M needs 18.4 GB but only 14.4 GB is addressable, so about 23% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 28 tokens per second — usable for batch work, painful for chat.
Mamba-2 + MoE hybrid built for the execution layer of agents: only 6 attention blocks, so the KV cache is almost free. Weights, data and recipe all open.
What hardware do I need for Nemotron 3.5 Lightning 30B-A3B? →
The VRAM budget
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 31.3 GB | 31.9 GB | — | ~7.5 | −0.1% ppl | 17.5 GB over |
| Q6_K | 24.1 GB | 24.8 GB | — | ~12 | −0.4% ppl | 10.4 GB over |
| Q5_K_M | 20.9 GB | 21.5 GB | — | ~17 | −0.8% ppl | 7.1 GB over |
| Q4_K_M | 17.8 GB | 18.4 GB | — | ~28 | −1.9% ppl | 4.0 GB over |
| Q3_K_M | 14.4 GB | 15.0 GB | — | ~93 | −5.4% ppl | 0.6 GB over |
| Q2_K | 12.3 GB | 13.0 GB | 251K | 179 | −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 6 of its 52 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-3.5-Lightning-30B-A3B-BF16:Q4_K_M \
-c 8192 -ngl 40
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.