Only with CPU offload
Nemotron 3.5 Lightning 30B-A3B at Q4_K_M needs 18.4 GB but only 7.0 GB is addressable, so about 64% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 11 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.
Hardware requirements What hardware do I need to run Nemotron 3.5 Lightning 30B-A3B? →At Q4_K_M with an 8K context: a 24 GB GPU (RTX 4090) or a 32 GB Mac (M4). Minimum, recommended, high-quality and long-context tiers, with all 114 devices checked at every quantisation.
The VRAM budget
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 31.3 GB | 31.9 GB | — | ~5.4 | −0.1% ppl | 24.9 GB over |
| Q6_K | 24.1 GB | 24.8 GB | — | ~7.4 | −0.4% ppl | 17.8 GB over |
| Q5_K_M | 20.9 GB | 21.5 GB | — | ~9.0 | −0.8% ppl | 14.5 GB over |
| Q4_K_M | 17.8 GB | 18.4 GB | — | ~11 | −1.9% ppl | 11.4 GB over |
| Q3_K_M | 14.4 GB | 15.0 GB | — | ~16 | −5.4% ppl | 8.0 GB over |
| Q2_K | 12.3 GB | 13.0 GB | — | ~20 | −15% ppl | 6.0 GB over |
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 18
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.
Related models on a GeForce RTX 3070
| Model | Total | Tok/s est. | |
|---|---|---|---|
| Nemotron 3 Super 120B-A12B 124B MoE · same family |
Q4_K_M · 70.4 GB | — | Won't fit |
| Gemma 4 31B 31.3B · similar size |
Q4_K_M · 20.2 GB | ~2.6 | CPU offload |
| GLM-4.7-Flash 30B-A3B 31.2B MoE · similar size |
Q4_K_M · 18.6 GB | ~12 | CPU offload |
| Olmo 3.1 32B Instruct 32.2B · similar size |
Q4_K_M · 20.0 GB | ~2.7 | CPU offload |
| Qwen3 30B-A3B 30.5B MoE · similar size |
Q4_K_M · 18.5 GB | ~11 | CPU offload |
| Qwen3 Coder 30B-A3B 30.5B MoE · similar size |
Q4_K_M · 18.5 GB | ~11 | CPU offload |
| Qwen3 32B 32.8B · similar size |
Q4_K_M · 21.0 GB | — | Won't fit |
Nemotron 3.5 Lightning 30B-A3B on other hardware
| Device | Total | Tok/s est. | |
|---|---|---|---|
| GeForce RTX 4060 Laptop 7.0 GB usable · 256 GB/s |
18.4 GB | ~11 | CPU offload |
| GeForce RTX 3060 12 GB 10.6 GB usable · 360 GB/s |
18.4 GB | ~15 | CPU offload |
| M4 · 16 GB 10.7 GB usable · 120 GB/s |
18.4 GB | — | Won't fit |
| GeForce RTX 4090 22.4 GB usable · 1008 GB/s |
18.4 GB | 130 | Runs great |
| Radeon RX 7900 XTX 22.4 GB usable · 960 GB/s |
18.4 GB | 124 | Runs great |
| M4 · 32 GB 24.0 GB usable · 120 GB/s |
18.4 GB | 17 | Runs great |
| M3 Max · 64 GB 48.0 GB usable · 400 GB/s |
18.4 GB | 58 | Runs great |