Can an RTX 3090 run Llama 3.3 70B Instruct?
Barely
It only loads at IQ2_XXS, which compresses the weights far enough to visibly degrade the model, so we would not call that running it.
In practice
It loads, and we would not call that running it. The only quantization that fits the GeForce RTX 3090 is IQ2_XXS, which compresses the weights far enough that Llama 3.3 70B Instruct loses a visible amount of its ability, so we do not recommend it. At Q4_K_M it needs 42.12 GB against the 23.2 GB available.
Running it properly means a split with system RAM: at Q4_K_M about 43 of 80 layers fit on the card, and the estimate falls to 1.5 tokens per second. A smaller model at a proper quantization is a better answer than a larger one squeezed in, and the next section lists the ones that fit.
The numbers
Llama 3.3 70B Instruct has 70.55 billion parameters across 80 layers. The GeForce RTX 3090 has 24 GB of VRAM at 936 GB/s, of which about 23.2 GB is left for a model once the operating system has taken its share.
| Quantization | Total needed | Fits | Speed |
|---|---|---|---|
| Q8_0 | 71.81 GB | partial | 0.6 t/s |
| Q6_K | 56.12 GB | partial | 0.9 t/s |
| Q5_K_M | 49.08 GB | partial | 1.1 t/s |
| Q4_K_M | 42.12 GB | partial | 1.5 t/s |
| IQ4_XS | 37.43 GB | partial | 2.0 t/s |
| Q3_K_M | 34.68 GB | partial | 2.4 t/s |
| IQ3_XXS degraded | 27.8 GB | partial | 5.2 t/s |
| Q2_K degraded | 30.15 GB | partial | 3.7 t/s |
| IQ2_XXS degraded | 19.71 GB | yes | 35.9 t/s |
1 of the 9 levels fit entirely. Levels marked degraded are listed for completeness, not as advice. Rows marked partial or experts in RAM run with part of the model in system RAM, at the speed shown.
What to run instead
A smaller model from the same family is the straightforward answer. These fit on an RTX 3090 at a quantization worth using:
- Llama 3.1 8B Instruct, 8.03B at Q8_0, 8.83 GB, about 81.7 tokens/s
- Llama 3.2 3B Instruct, 3.21B at Q8_0, 4.42 GB, about 170 tokens/s
- Llama 3.2 1B Instruct, 1.24B at Q8_0, 1.82 GB, about 466 tokens/s
Shortening the context will not rescue this one: at 4k it still wants 40.87 GB against 42.12 GB at 8k, because the weights, not the cache, are what fill the card.
Models of this size are what a second GPU or a large-memory machine is for. The multi-GPU guide shows what 2 cards add, and the Mac and PC comparison shows where unified memory holds a model that no single card can.
Or the hardware that does run it
The cheapest device we track that runs Llama 3.3 70B Instruct at a quantization worth using is the RTX A6000, around $3,500 used , running it at IQ4_XS at about 15.3 tokens per second. Check the current price (affiliate link, indicative price reviewed 2026-09-01). Compare every option.
What else an RTX 3090 runs well
For comparison, these are the other popular models that fit this device at a sound quantization, best pairings first.
| Model | Quantization | Memory | Speed |
|---|---|---|---|
| Qwen3.5 27B | Q6_K | 21.33 GB | 33.0 t/s |
| Gemma 4 31B Instruct | Q4_K_M | 21.31 GB | 33.0 t/s |
| Qwen3 32B | Q4_K_M | 20.42 GB | 34.5 t/s |
| Gemma 4 12B Instruct | Q8_0 | 13.55 GB | 52.4 t/s |
The same model on nearby hardware
Memory decides whether Llama 3.3 70B Instruct loads and bandwidth decides how quickly it answers. Here is what a step down, a card with the same memory and a step up do with it.
| Device | Memory | Bandwidth | Result |
|---|---|---|---|
| GeForce RTX 5080 | 16 GB | 960 GB/s | partial, 1.1 t/s |
| GeForce RTX 4090 | 24 GB | 1008 GB/s | does not fit |
| GeForce RTX 5090 | 32 GB | 1792 GB/s | does not fit |
Questions people ask
How much VRAM does Llama 3.3 70B Instruct need on RTX 3090?
At Q4_K_M with an 8k context window it needs 42.12 GB in total: 39.08 GB of weights, 2.5 GB of KV cache and 0.54 GB of runtime overhead. The GeForce RTX 3090 makes about 23.2 GB available once the operating system has taken its share.
What should I run instead of Llama 3.3 70B Instruct on RTX 3090?
Llama 3.1 8B Instruct fits at Q8_0 and runs at about 81.7 tokens per second. Smaller models from the same family are the usual answer.
Keep going
Everything the GeForce RTX 3090 runs, the full breakdown for Llama 3.3 70B Instruct, how the numbers are made in the methodology, or a different pairing in the calculator.