← All models
NVIDIA · mixture of experts · OpenMDW 1.1
What runs Nemotron 3.5 Lightning 30B-A3B
Hybrid Mamba/MoE model with 30B total and 3B active parameters. Very small KV growth makes long contexts cheap on 24 GB GPUs.
- Parameters
- 30B, 3B active
- 4-bit file
- 23.7 GB
- Q4_K_M
- Needs at 8k context
- 25.3 GB
- context memory estimated
- Longest context
- 262,144 tokens
Context memory: 6,144 bytes per token for the full-attention layers only. Mamba-2 + MoE + attention hybrid: only 6 of 52 layers are attention (2 x 6 x 2 KV heads x 128 x 2 = 6,144 bytes/token); Mamba layers hold a fixed-size state. Card gives no per-token figure. How context memory works.
Measured on this hardware
Machines that hold the 4-bit file with 8,192 tokens of context and have published runs on the same chip or cards, fastest first. Median of single-user runs with plain decoding.
| Machine | Fit | Price | Generation |
|---|---|---|---|
| Apple Mac Studio (M5 Ultra, 64-core GPU)Apple unified memory | Fitsassumed limitfrom 96 GB | ~$5,499–$9,499 | 196 tok/smedian of 1 run |
| Apple Mac Studio (M5 Max, 40-core GPU)Apple unified memory | Fitsassumed limitfrom 48 GB | ~$3,099–$5,099 | 142 tok/smedian of 2 runs |
| Apple MacBook Pro 14-inch (M5 Max, 40-core GPU)Apple unified memory | Fitsassumed limitfrom 48 GB | ~$4,699–$6,699 | 142 tok/smedian of 2 runs |
| Apple MacBook Pro 16-inch (M5 Max, 40-core GPU)Apple unified memory | Fitsassumed limitfrom 48 GB | ~$4,999–$6,999 | 142 tok/smedian of 2 runs |
| Apple Mac mini (M5 Pro, 20-core GPU)Apple unified memory | Fitsassumed limitfrom 48 GB | ~$1,899–$2,899 | 106 tok/smedian of 2 runs |
Estimated, not measured
These fit but nobody has published a run we could match. The figure is memory bandwidth divided by the bytes read per token, scaled by the share of that ceiling measured runs reach. How estimates work.
| Machine | Fit | Price | Generation |
|---|---|---|---|
| Dell Pro Precision with GB300NVIDIA workstation | Fits | ~$159,283 | ~1,313–1,676 tok/sestimate |
| Dell Pro Max Tower T2 with NVIDIA RTX PRO 6000 BlackwellNVIDIA workstation | Fits | ~$13,909 | ~175–256 tok/sestimate |
| 1x RTX 5090Single GPU | Fits | ~$3,234 + unpriced parts | ~175–256 tok/sestimate |
| 2x RTX 5090Multi-GPU | Fits | ~$4,109 + unpriced parts | ~175–256 tok/sestimate |
| 1x RTX PRO 6000 BlackwellSingle GPU | Fits | ~$20,833 | ~175–256 tok/sestimate |
| RTX 5090 + EPYC 9355P with 768 GB DDR5 (MoE offload)Hybrid GPU + RAM | Fits | ~$3,763 + unpriced parts | ~175–256 tok/sestimate |
| Apple Mac Studio (M5 Ultra, 80-core GPU)Apple unified memory | Fitsassumed limitfrom 96 GB | ~$6,799–$10,799 | ~157–279 tok/sestimate |
| 4x Instinct MI50 32GB (used) on EPYC 7003Multi-GPU | Fits | ~$1,998 + unpriced parts | ~100–146 tok/sestimate |
| 2x Radeon RX 7900 XTXMulti-GPU | Fits | ~$5,637 | ~94–137 tok/sestimate |
| 2x RTX 3090 (used)Multi-GPU | Fits | ~$3,639 + unpriced parts | ~91–133 tok/sestimate |
| 4x RTX 3090 (used) on EPYC 7003, open frameMulti-GPU | Fits | ~$2,178 + unpriced parts | ~91–133 tok/sestimate |
| Apple Mac Studio (M5 Max, 32-core GPU)Apple unified memory | Fits with raised limitassumed limit | ~$2,499 | ~60–107 tok/sestimate |
| Apple MacBook Pro 14-inch (M5 Max, 32-core GPU)Apple unified memory | Fits with raised limitassumed limit | ~$4,099 | ~60–107 tok/sestimate |
| Apple MacBook Pro 16-inch (M5 Max, 32-core GPU)Apple unified memory | Fits with raised limitassumed limit | ~$4,399 | ~60–107 tok/sestimate |
| RTX 3090 (used) + 128 GB DDR5 on AM5 (MoE offload)Hybrid GPU + RAM | GPU + RAM offloadassumed limit | ~$4,833 + unpriced parts | ~64–93 tok/sestimate, part in RAM |
| tiny corp tinybox red flashGPU desktop (prebuilt) | Fits | ~$22,000 | ~63–91 tok/sestimate |
| Framework Desktop (Ryzen AI Max+ PRO 495)AMD unified memory | Fitsassumed limit | ~$6,799 | ~50–64 tok/sestimate |
| NVIDIA DGX SparkNVIDIA GB10 | Fitsassumed limit | ~$4,699 | ~50–64 tok/sestimate |
| Dell Pro Precision with GB10NVIDIA GB10 | Fitsassumed limit | ~$5,523 | ~50–64 tok/sestimate |
| 2x Intel Arc Pro B60Multi-GPU | Fits | ~$4,367 | ~45–65 tok/sestimate |
| Framework Desktop (Ryzen AI Max+ 395)AMD unified memory | Fitsassumed limitfrom 64 GB | ~$1,959–$3,449 | ~47–60 tok/sestimate |
| GMKtec EVO-X2 (Ryzen AI Max+ 395)AMD unified memory | Fitsassumed limitfrom 64 GB | ~$2,600–$4,000 | ~47–60 tok/sestimate |
| Beelink GTR9 Pro (Ryzen AI Max+ 395)AMD unified memory | Fits | ~$4,699 | ~47–60 tok/sestimate |
| Minisforum MS-S1 MAX (Ryzen AI Max+ 395)AMD unified memory | Fits | ~$4,899 | ~47–60 tok/sestimate |
| Corsair AI Workstation 300 (Ryzen AI Max+ 395)AMD unified memory | Fits | ~$4,800 | ~47–60 tok/sestimate |
| HP Z2 Mini G1a (Ryzen AI Max+ PRO 395)AMD unified memory | Fits | ~$4,999 | ~47–60 tok/sestimate |
| Apple Mac mini (M5 Pro, 16-core GPU)Apple unified memory | Fitsassumed limitfrom 48 GB | ~$1,699–$2,699 | ~40–71 tok/sestimate |
Fits, no speed figure
Too large for
Framework Desktop (Ryzen AI Max 385), 1x RTX 3090 (used). A smaller quant may fit: try the calculator.
Files
| Quant | Format | Size | Repo |
|---|---|---|---|
| 4bit | MLX | 16.6 GB | mlx-community/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-4bit |
| Q4_K_M | GGUF | 23.7 GB | bartowski/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF |
| Q5_K_M | GGUF | 25.1 GB | bartowski/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF |
| 8bit | MLX | 31.3 GB | mlx-community/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-8bit |
| Q6_K | GGUF | 31.9 GB | bartowski/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF |
| Q8_0 | GGUF | 32.6 GB | bartowski/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF |
Sources
- huggingface.co: model card: params, context, license, attention layout, description
- huggingface.co: downloads (last 30 days), license tag, safetensors param count
- huggingface.co: layer count, KV heads and head size
- huggingface.co: GGUF file sizes (shards summed)
- huggingface.co: MLX 4bit safetensors sizes
- huggingface.co: MLX 8bit safetensors sizes