Llama · 32B to 80B
Llama 3.3 70B Instruct
70.55 billion parameters across 80 layers. At four-bit precision the weights alone come to 39.67 GB, before any conversation is loaded. Grouped-query attention keeps the cache small, so long contexts cost less here than on models of the same size.
Memory needed, by quantisation
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.
| Quantisation | File size | KV cache | Total VRAM | Quality | What it costs you |
|---|---|---|---|---|---|
| Q8_0 | 69.81 GB | 2.5 GB | 73.41 GB | 99% | Lossless in practice. Use it when the memory is there. |
| Q6_K | 53.88 GB | 2.5 GB | 57.48 GB | 98% | Very close to Q8 for two thirds of the size. |
| Q5_K_M | 46.73 GB | 2.5 GB | 50.33 GB | 96% | The quality-first choice when Q6 will not fit. |
| Q4_K_M | 39.67 GB | 2.5 GB | 43.27 GB | 93% | The default. Best size-to-quality ratio for local use. |
| IQ4_XS | 34.91 GB | 2.5 GB | 38.51 GB | 91% | Importance-matrix 4-bit. Q4_K_S quality, smaller file. |
| Q3_K_M | 32.11 GB | 2.5 GB | 35.71 GB | 86% | Degradation starts to show. A way to fit one size up. |
| IQ3_XXS | 25.13 GB | 2.5 GB | 28.73 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 | 1.25 GB | 41.77 GB | no | no | no |
| 8k | 2.5 GB | 43.27 GB | no | no | no |
| 16k | 5 GB | 46.27 GB | no | no | no |
| 32k | 10 GB | 52.27 GB | no | no | no |
| 64k | 20 GB | 64.27 GB | no | no | no |
| 128k | 40 GB | 88.27 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 70B Instruct
29 of 118 consumer devices run it at a quantisation worth using, at an 8k context window. Another 20 can load it only by compressing the weights far enough to damage the model, marked with a warning below.
| Device | Memory | Verdict | Quantisation | Used | Speed |
|---|---|---|---|---|---|
| RTX PRO 6000 Blackwell NVIDIA | 96 GB | Runs well | Q8_0 | 73.41 GB | 20.3 t/s |
| RTX 6000 Ada Generation NVIDIA | 48 GB | Runs well | Q4_K_M | 43.27 GB | 18.7 t/s |
| RTX A6000 NVIDIA | 48 GB | Runs well | Q4_K_M | 43.27 GB | 14.9 t/s |
| Radeon PRO W7900 AMD | 48 GB | Runs well | Q4_K_M | 43.27 GB | 14.8 t/s |
| Apple M1 Ultra 64GB Apple | 64 GB | Runs well | Q4_K_M | 42.97 GB | 14.8 t/s |
| Apple M2 Ultra 64GB Apple | 64 GB | Runs well | Q4_K_M | 42.97 GB | 14.8 t/s |
| Apple M3 Ultra 96GB Apple | 96 GB | Runs well | Q6_K | 57.18 GB | 11.3 t/s |
| Apple M3 Ultra 256GB Apple | 256 GB | Runs well | Q6_K | 57.18 GB | 11.3 t/s |
| Apple M3 Ultra 512GB Apple | 512 GB | Runs well | Q6_K | 57.18 GB | 11.3 t/s |
| Apple M1 Ultra 128GB Apple | 128 GB | Runs well | Q6_K | 57.18 GB | 11.1 t/s |
| Apple M2 Ultra 128GB Apple | 128 GB | Runs well | Q6_K | 57.18 GB | 11.1 t/s |
| Apple M2 Ultra 192GB Apple | 192 GB | Runs well | Q6_K | 57.18 GB | 11.1 t/s |
| Apple M4 Max 64GB Apple | 64 GB | Runs well | Q4_K_M | 42.97 GB | 10.1 t/s |
| Apple M4 Max 128GB Apple | 128 GB | Runs well | Q4_K_M | 42.97 GB | 10.1 t/s |
| GeForce RTX 5090 NVIDIA | 32 GB | Runs great | IQ3_XXS ! | 28.73 GB | 53.2 t/s |
| Apple M4 Max 48GB Apple | 48 GB | Just fits | Q3_K_M | 35.41 GB | 12.3 t/s |
| Apple M1 Max 64GB Apple | 64 GB | Fits, but slow | Q4_K_M | 42.97 GB | 7.4 t/s |
| Apple M2 Max 64GB Apple | 64 GB | Fits, but slow | Q4_K_M | 42.97 GB | 7.4 t/s |
| Apple M3 Max 64GB Apple | 64 GB | Fits, but slow | Q4_K_M | 42.97 GB | 7.4 t/s |
| Apple M2 Max 96GB Apple | 96 GB | Fits, but slow | Q6_K | 57.18 GB | 5.5 t/s |
| Apple M3 Max 96GB Apple | 96 GB | Fits, but slow | Q6_K | 57.18 GB | 5.5 t/s |
| Apple M3 Max 48GB Apple | 48 GB | Just fits | Q3_K_M | 35.41 GB | 9 t/s |
| Apple M3 Max 128GB Apple | 128 GB | Fits, but slow | Q8_0 | 73.11 GB | 4.3 t/s |
| Apple M4 Pro 64GB Apple | 64 GB | Fits, but slow | Q4_K_M | 42.97 GB | 5 t/s |
| NVIDIA DGX Spark 128GB NVIDIA | 128 GB | Fits, but slow | Q8_0 | 73.41 GB | 3.1 t/s |
| Apple M4 Pro 48GB Apple | 48 GB | Just fits | Q3_K_M | 35.41 GB | 6.2 t/s |
| Jetson AGX Orin 64GB NVIDIA | 64 GB | Fits, but slow | Q4_K_M | 43.27 GB | 4 t/s |
| Ryzen AI Max+ 395 64GB AMD | 64 GB | Fits, but slow | Q4_K_M | 43.27 GB | 3.9 t/s |
| Ryzen AI Max+ 395 96GB AMD | 96 GB | Fits, but slow | Q6_K | 57.48 GB | 3 t/s |
| Ryzen AI Max+ 395 128GB AMD | 128 GB | Fits, but slow | Q8_0 | 73.41 GB | 2.3 t/s |
| GeForce RTX 4090 NVIDIA | 24 GB | Runs great | IQ2_XXS ! | 20.52 GB | 42.6 t/s |
| GeForce RTX 3090 Ti NVIDIA | 24 GB | Runs great | IQ2_XXS ! | 20.52 GB | 42.6 t/s |
| GeForce RTX 3090 NVIDIA | 24 GB | Runs great | IQ2_XXS ! | 20.52 GB | 39.5 t/s |
| GeForce RTX 5090 Laptop NVIDIA | 24 GB | Runs great | IQ2_XXS ! | 20.52 GB | 37.8 t/s |
| Radeon RX 7900 XTX AMD | 24 GB | Runs great | IQ2_XXS ! | 20.52 GB | 35.6 t/s |
| RTX A5000 NVIDIA | 24 GB | Runs great | IQ2_XXS ! | 20.52 GB | 32.4 t/s |
| Apple M4 Max 36GB Apple | 36 GB | Runs well | IQ2_XXS ! | 20.22 GB | 21.9 t/s |
| Apple M1 Max 32GB Apple | 32 GB | Runs well | IQ2_XXS ! | 20.22 GB | 16.1 t/s |
| Apple M2 Max 32GB Apple | 32 GB | Runs well | IQ2_XXS ! | 20.22 GB | 16.1 t/s |
| Apple M3 Max 36GB Apple | 36 GB | Runs well | IQ2_XXS ! | 20.22 GB | 16.1 t/s |
| Ryzen AI Max+ 395 32GB AMD | 32 GB | Fits, but slow | IQ2_XXS ! | 20.52 GB | 8.6 t/s |
| Jetson AGX Orin 32GB NVIDIA | 32 GB | Fits, but slow | IQ2_XXS ! | 20.52 GB | 8.6 t/s |
| Apple M1 Pro 32GB Apple | 32 GB | Fits, but slow | IQ2_XXS ! | 20.22 GB | 8 t/s |
| Apple M2 Pro 32GB Apple | 32 GB | Fits, but slow | IQ2_XXS ! | 20.22 GB | 8 t/s |
| Apple M5 32GB Apple | 32 GB | Fits, but slow | IQ2_XXS ! | 20.22 GB | 6.1 t/s |
| Apple M3 Pro 36GB Apple | 36 GB | Fits, but slow | IQ2_XXS ! | 20.22 GB | 6 t/s |
| Apple M4 32GB Apple | 32 GB | Fits, but slow | IQ2_XXS ! | 20.22 GB | 4.8 t/s |
| Intel Core Ultra 9 288V 32GB Intel | 32 GB | Fits, but slow | IQ2_XXS ! | 20.52 GB | 4.6 t/s |
| Ryzen AI 9 HX 370 32GB AMD | 32 GB | Fits, but slow | IQ2_XXS ! | 20.52 GB | 4.3 t/s |
| Radeon RX 7900 XT AMD | 20 GB | Partial offload 34/80 | — | 43.27 GB | 1.3 t/s |
| GeForce RTX 5080 NVIDIA | 16 GB | Partial offload 26/80 | — | 43.27 GB | 1.1 t/s |
| GeForce RTX 5070 Ti NVIDIA | 16 GB | Partial offload 26/80 | — | 43.27 GB | 1.1 t/s |
| GeForce RTX 5060 Ti 16GB NVIDIA | 16 GB | Partial offload 26/80 | — | 43.27 GB | 1.1 t/s |
| GeForce RTX 4080 SUPER NVIDIA | 16 GB | Partial offload 26/80 | — | 43.27 GB | 1.1 t/s |
| GeForce RTX 4080 NVIDIA | 16 GB | Partial offload 26/80 | — | 43.27 GB | 1.1 t/s |
| GeForce RTX 4070 Ti SUPER NVIDIA | 16 GB | Partial offload 26/80 | — | 43.27 GB | 1.1 t/s |
| GeForce RTX 4090 Laptop NVIDIA | 16 GB | Partial offload 26/80 | — | 43.27 GB | 1.1 t/s |
| GeForce RTX 5080 Laptop NVIDIA | 16 GB | Partial offload 26/80 | — | 43.27 GB | 1.1 t/s |
| RTX A4000 NVIDIA | 16 GB | Partial offload 26/80 | — | 43.27 GB | 1.1 t/s |
| Radeon RX 9070 XT AMD | 16 GB | Partial offload 26/80 | — | 43.27 GB | 1.1 t/s |
What to buy to run Llama 3.3 70B Instruct
The cheapest hardware that runs it at a quantisation worth using and a speed you would not resent, at an 8k context window.
Cheapest that works
Apple M1 Ultra 64GB
$2,400 whole machine
IQ4_XS · uses 38.21 GB of its 64 GB · about 16.7 tokens/s
Best value
Apple M1 Ultra 128GB
$3,400 whole machine
IQ4_XS · about 16.7 tokens/s · 800 GB/s
If it just has to run
Ryzen AI Max+ 395 64GB
$1,699 whole machine
IQ4_XS · about 4.4 tokens/s. Cheaper, and slower to answer.
| Device | Price | Memory | Runs it at | Speed | |
|---|---|---|---|---|---|
| Apple M1 Ultra 64GB Apple · whole machine | $2,400 | 64 GB | IQ4_XS | 16.7 t/s | Check price |
| Apple M1 Ultra 128GB Apple · whole machine | $3,400 | 128 GB | IQ4_XS | 16.7 t/s | Check price |
| RTX A6000 NVIDIA | $3,500 used | 48 GB | IQ4_XS | 16.8 t/s | Check price |
| Apple M3 Ultra 256GB Apple · whole machine | $5,599 | 256 GB | Q4_K_M | 15.1 t/s | Check price |
| RTX 6000 Ada Generation NVIDIA | $6,800 | 48 GB | Q4_K_M | 18.7 t/s | Check price |
| RTX PRO 6000 Blackwell NVIDIA | $8,500 | 96 GB | Q8_0 | 20.3 t/s | Check price |
| Apple M3 Ultra 512GB Apple · whole machine | $9,499 | 512 GB | Q4_K_M | 15.1 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.
These numbers also cover
Fine-tunes share their base model's architecture, so they need exactly the same memory. If you are looking for one of these, the figures above apply unchanged.
meta-llama/Llama-3.1-70B-Instructnvidia/Llama-3.1-Nemotron-70B-Instruct-HFdeepseek-ai/DeepSeek-R1-Distill-Llama-70Blatam-gpt/Llama-3.1-70B-LatamGPT-SFT-1.0
Where it sits in the catalogue
Source: meta-llama/Llama-3.3-70B-Instruct on Hugging Face. Downloaded 549k times in the last month. Published 2024-11-26. Architecture figures are read from the repository's own configuration file, so they move when the model does.
Llama 3 3 Nemotron Super 49B V1
49.87B · 128k context
Llama 3.1 8B Instruct
8.03B · 128k context
Llama 3 Taiwan 8B Instruct
8.03B · 8k context
CodeLlama 7B
6.74B · 16k context
Llama 3.2 3B Instruct
3.21B · 128k context
Llama 3.2 1B Instruct
1.24B · 128k context
TinyLlama 1.1B Chat V1.0
1.1B · 2k context
Llama 160m
0.16B · 2k context