Llama · 32B to 80B
Llama 3 3 Nemotron Super 49B V1
49.87 billion parameters across 80 layers. At four-bit precision the weights alone come to 28.04 GB, before any conversation is loaded. This one is cache-hungry: a 32k window adds 80.0 GB on top, so context length matters more here than the model size suggests.
About Llama 3 3 Nemotron Super 49B V1
Llama 3 3 Nemotron Super 49B V1 is a dense model with 49.87B parameters, built from 80 layers with a hidden size of 8192. Its repository was published on 2025-03-16. The license listed on Hugging Face is other. It keeps one KV pair for every attention head, so its cache is large for its size.
Memory needed, by quantization
At an 8k context window. Quality is our rough ranking of how much the compression costs you: anything at or above Q5 is hard to tell apart from the original in normal use.
| Quantization | File size | KV cache | Total VRAM | Quality | What it costs you |
|---|---|---|---|---|---|
| Q8_0 | 49.35 GB | 20 GB | 68.85 GB | 99% | Lossless in practice. Use it when the memory is there. |
| Q6_K | 38.08 GB | 20 GB | 57.82 GB | 98% | Very close to Q8 for two thirds of the size. |
| Q5_K_M | 33.03 GB | 20 GB | 52.88 GB | 96% | The quality-first choice when Q6 will not fit. |
| Q4_K_M | 28.04 GB | 20 GB | 47.99 GB | 93% | The default. Best size-to-quality ratio for local use. |
| IQ4_XS | 24.67 GB | 20 GB | 44.7 GB | 91% | Importance-matrix 4-bit. Q4_K_S quality, smaller file. |
| Q3_K_M | 22.7 GB | 20 GB | 42.76 GB | 86% | Degradation starts to show. A way to fit one size up. |
| IQ3_XXS | 17.77 GB | 20 GB | 37.93 GB | 79% | Aggressive. Only worth it on very large models. |
What a longer conversation costs
Same model at Q4_K_M, only the context window changes. The cache grows in a straight line with every token in the window.
| Context | KV cache | Total VRAM | Fits in 8 GB | Fits in 12 GB | Fits in 24 GB |
|---|---|---|---|---|---|
| 4k | 10 GB | 37.99 GB | no | no | no |
| 8k | 20 GB | 47.99 GB | no | no | no |
| 16k | 40 GB | 67.99 GB | no | no | no |
| 32k | 80 GB | 107.99 GB | no | no | no |
| 64k | 160 GB | 187.99 GB | no | no | no |
| 128k | 320 GB | 347.99 GB | no | no | no |
The "fits" columns allow for the roughly 0.8 GB Windows keeps for the desktop.
Which hardware runs Llama 3 3 Nemotron Super 49B V1
17 of 118 consumer devices run it at a quantization worth using, at an 8k context window.
| Device | Memory | Verdict | Quantization | Used | Speed |
|---|---|---|---|---|---|
| RTX PRO 6000 Blackwell NVIDIA | 96 GB | Runs well | Q8_0 | 68.85 GB | 19.3 t/s |
| RTX 6000 Ada Generation NVIDIA | 48 GB | Runs well | IQ4_XS | 44.7 GB | 16 t/s |
| Apple M3 Ultra 96GB Apple | 96 GB | Runs well | Q5_K_M | 53.55 GB | 10.6 t/s |
| Apple M3 Ultra 256GB Apple | 256 GB | Runs well | Q5_K_M | 53.55 GB | 10.6 t/s |
| Apple M3 Ultra 512GB Apple | 512 GB | Runs well | Q5_K_M | 53.55 GB | 10.6 t/s |
| Apple M1 Ultra 128GB Apple | 128 GB | Runs well | Q5_K_M | 53.55 GB | 10.4 t/s |
| Apple M2 Ultra 128GB Apple | 128 GB | Runs well | Q5_K_M | 53.55 GB | 10.4 t/s |
| Apple M2 Ultra 192GB Apple | 192 GB | Runs well | Q5_K_M | 53.55 GB | 10.4 t/s |
| RTX A6000 NVIDIA | 48 GB | Runs well | IQ4_XS | 44.7 GB | 12.8 t/s |
| Radeon PRO W7900 AMD | 48 GB | Runs well | IQ4_XS | 44.7 GB | 12.2 t/s |
| Apple M4 Max 128GB Apple | 128 GB | Fits, but slow | Q8_0 | 69.86 GB | 5.4 t/s |
| Apple M2 Max 96GB Apple | 96 GB | Fits, but slow | Q8_0 | 69.86 GB | 4 t/s |
| Apple M3 Max 96GB Apple | 96 GB | Fits, but slow | Q8_0 | 69.86 GB | 4 t/s |
| Apple M3 Max 128GB Apple | 128 GB | Fits, but slow | Q8_0 | 69.86 GB | 4 t/s |
| NVIDIA DGX Spark 128GB NVIDIA | 128 GB | Fits, but slow | Q8_0 | 69.89 GB | 2.9 t/s |
| Ryzen AI Max+ 395 96GB AMD | 96 GB | Fits, but slow | Q8_0 | 69.89 GB | 2.1 t/s |
| Ryzen AI Max+ 395 128GB AMD | 128 GB | Fits, but slow | Q8_0 | 69.89 GB | 2.1 t/s |
| Apple M4 Max 36GB Apple | 36 GB | Just fits | n/a ! | 32.47 GB | 11.7 t/s |
| Apple M1 Ultra 64GB Apple | 64 GB | Just fits | n/a | 53.55 GB | 10.4 t/s |
| Apple M2 Ultra 64GB Apple | 64 GB | Just fits | n/a | 53.55 GB | 10.4 t/s |
| Apple M4 Max 48GB Apple | 48 GB | Just fits | n/a | 43.21 GB | 8.8 t/s |
| Apple M3 Max 36GB Apple | 36 GB | Just fits | n/a ! | 32.47 GB | 8.6 t/s |
| Apple M4 Max 64GB Apple | 64 GB | Just fits | n/a | 58.6 GB | 6.5 t/s |
| Apple M3 Max 48GB Apple | 48 GB | Just fits | n/a | 43.21 GB | 6.4 t/s |
| Apple M1 Max 64GB Apple | 64 GB | Just fits | n/a | 58.6 GB | 4.7 t/s |
Show 35 more devices
| Apple M2 Max 64GB Apple | 64 GB | Just fits | n/a | 58.6 GB | 4.7 t/s |
| Apple M3 Max 64GB Apple | 64 GB | Just fits | n/a | 58.6 GB | 4.7 t/s |
| Apple M4 Pro 48GB Apple | 48 GB | Just fits | n/a | 43.21 GB | 4.4 t/s |
| Apple M3 Pro 36GB Apple | 36 GB | Just fits | n/a ! | 32.47 GB | 3.2 t/s |
| Apple M4 Pro 64GB Apple | 64 GB | Just fits | n/a | 58.6 GB | 3.2 t/s |
| Jetson AGX Orin 64GB NVIDIA | 64 GB | Just fits | n/a | 58.63 GB | 2.6 t/s |
| Ryzen AI Max+ 395 64GB AMD | 64 GB | Just fits | n/a | 58.63 GB | 2.5 t/s |
| GeForce RTX 5090 NVIDIA | 32 GB | Partial offload 51/80 | n/a | 47.99 GB | 1.7 t/s |
| GeForce RTX 4090 NVIDIA | 24 GB | Partial offload 38/80 | n/a | 47.99 GB | 1.2 t/s |
| GeForce RTX 3090 Ti NVIDIA | 24 GB | Partial offload 38/80 | n/a | 47.99 GB | 1.2 t/s |
| GeForce RTX 3090 NVIDIA | 24 GB | Partial offload 38/80 | n/a | 47.99 GB | 1.2 t/s |
| GeForce RTX 5090 Laptop NVIDIA | 24 GB | Partial offload 38/80 | n/a | 47.99 GB | 1.2 t/s |
| RTX A5000 NVIDIA | 24 GB | Partial offload 38/80 | n/a | 47.99 GB | 1.1 t/s |
| Radeon RX 7900 XTX AMD | 24 GB | Partial offload 38/80 | n/a | 47.99 GB | 1.1 t/s |
| Radeon RX 7900 XT AMD | 20 GB | Partial offload 31/80 | n/a | 47.99 GB | 1 t/s |
| GeForce RTX 5080 NVIDIA | 16 GB | Partial offload 24/80 | n/a | 47.99 GB | 0.9 t/s |
| GeForce RTX 5070 Ti NVIDIA | 16 GB | Partial offload 24/80 | n/a | 47.99 GB | 0.9 t/s |
| GeForce RTX 5060 Ti 16GB NVIDIA | 16 GB | Partial offload 24/80 | n/a | 47.99 GB | 0.9 t/s |
| GeForce RTX 4080 SUPER NVIDIA | 16 GB | Partial offload 24/80 | n/a | 47.99 GB | 0.9 t/s |
| GeForce RTX 4080 NVIDIA | 16 GB | Partial offload 24/80 | n/a | 47.99 GB | 0.9 t/s |
| GeForce RTX 4070 Ti SUPER NVIDIA | 16 GB | Partial offload 24/80 | n/a | 47.99 GB | 0.9 t/s |
| GeForce RTX 4090 Laptop NVIDIA | 16 GB | Partial offload 24/80 | n/a | 47.99 GB | 0.9 t/s |
| GeForce RTX 5080 Laptop NVIDIA | 16 GB | Partial offload 24/80 | n/a | 47.99 GB | 0.9 t/s |
| RTX A4000 NVIDIA | 16 GB | Partial offload 24/80 | n/a | 47.99 GB | 0.9 t/s |
| Radeon RX 9070 XT AMD | 16 GB | Partial offload 24/80 | n/a | 47.99 GB | 0.9 t/s |
| Radeon RX 9070 AMD | 16 GB | Partial offload 24/80 | n/a | 47.99 GB | 0.9 t/s |
| Radeon RX 7900 GRE AMD | 16 GB | Partial offload 24/80 | n/a | 47.99 GB | 0.9 t/s |
| Radeon RX 7800 XT AMD | 16 GB | Partial offload 24/80 | n/a | 47.99 GB | 0.9 t/s |
| Radeon RX 6900 XT AMD | 16 GB | Partial offload 24/80 | n/a | 47.99 GB | 0.9 t/s |
| Radeon RX 6800 AMD | 16 GB | Partial offload 24/80 | n/a | 47.99 GB | 0.9 t/s |
| Arc A770 16GB Intel | 16 GB | Partial offload 24/80 | n/a | 47.99 GB | 0.9 t/s |
| GeForce RTX 5070 NVIDIA | 12 GB | Partial offload 17/80 | n/a | 47.99 GB | 0.8 t/s |
| GeForce RTX 4070 Ti NVIDIA | 12 GB | Partial offload 17/80 | n/a | 47.99 GB | 0.8 t/s |
| GeForce RTX 4070 SUPER NVIDIA | 12 GB | Partial offload 17/80 | n/a | 47.99 GB | 0.8 t/s |
| GeForce RTX 4070 NVIDIA | 12 GB | Partial offload 17/80 | n/a | 47.99 GB | 0.8 t/s |
How it compares
Other models in the same family or of a similar size, on a 12 GB card and a 24 GB card, each at the best quantization that fits.
| Model | Weights at 4 bits | GeForce RTX 3060 12GB | GeForce RTX 4090 |
|---|---|---|---|
| Llama 3 3 Nemotron Super 49B V1 | 28.0 GB | does not fit well | does not fit well |
| Llama 3.3 70B Instruct | 39.7 GB | does not fit well | does not fit well |
| Llama 3.1 8B Instruct | 4.5 GB | 31.4 t/s at Q8_0 | 88 t/s at Q8_0 |
| Llama 3.2 3B Instruct | 1.8 GB | 65.4 t/s at Q8_0 | 183.1 t/s at Q8_0 |
| Qwen3 32B | 18.4 GB | does not fit well | 37.1 t/s at Q4_K_M |
| Qwen3.5 35B A3B | 20.2 GB | 55.7 t/s at Q4_K_M | 235.9 t/s at Q4_K_M |
What to buy to run Llama 3 3 Nemotron Super 49B V1
The cheapest hardware that runs it at a quantization worth using and a speed you would not resent, at an 8k context window.
Cheapest that works
RTX 6000 Ada Generation
$6,800
IQ4_XS · uses 44.7 GB of its 48 GB · about 16 tokens/s
Best value
RTX PRO 6000 Blackwell
$8,500
Q8_0 · about 19.3 tokens/s · 1792 GB/s
If it just has to run
Apple M4 Pro 48GB
$1,799 whole machine
Q3_K_M · about 4.4 tokens/s. Cheaper because it runs the model at a lower precision.
| Device | Price | Memory | Runs it at | Speed | |
|---|---|---|---|---|---|
| RTX 6000 Ada Generation NVIDIA | $6,800 | 48 GB | IQ4_XS | 16 t/s | Check price |
| RTX PRO 6000 Blackwell NVIDIA | $8,500 | 96 GB | Q8_0 | 19.3 t/s | Check price |
Indicative prices reviewed 2026-09-01; used prices are marketplace typical. Options that cost more than a cheaper one with no more memory and no more speed are hidden. Price links are Amazon affiliate links. Change the standard, the context window or the budget in the buying tool.
Where it sits in the catalog
Source: nvidia/Llama-3_3-Nemotron-Super-49B-v1 on Hugging Face. Downloaded 176k times in the last month. Published 2025-03-16. Architecture figures are read from the repository's own configuration file, so they move when the model does.
Questions people ask about Llama 3 3 Nemotron Super 49B V1
How much VRAM does Llama 3 3 Nemotron Super 49B V1 need?
At Q4_K_M with an 8k context window, Llama 3 3 Nemotron Super 49B V1 needs about 47.99 GB: 27.45 GB of weights, 20 GB of KV cache and 0.54 GB of overhead. At the highest-quality file it needs about 68.85 GB.
What is the cheapest GPU that runs Llama 3 3 Nemotron Super 49B V1?
The smallest consumer device in our catalog that fits it is the Apple M3 Ultra 96GB with 96 GB, running it at Q5_K_M at roughly 10.6 tokens per second.
How much does a longer context window add for Llama 3 3 Nemotron Super 49B V1?
On the memory side, a 32k window adds 80.0 GB of KV cache on top of the weights, and the cache grows in a straight line with every token. The longest window the model supports natively is 128k tokens.