Can an RTX 5090 run Llama 3.1 8B Instruct?
Yes
At Q8_0 it needs 8.83 GB of the 31.2 GB available and generates around 157 tokens per second at an 8k context window.
In practice
Llama 3.1 8B Instruct sits comfortably on RTX 5090. At Q8_0 it takes 8.83 GB of the 31.2 GB available, which leaves 22.37 GB spare, enough for the 8k window we assumed and for a window of up to about 128k tokens before memory runs out. Generation runs at about 157 tokens per second, well past the speed anyone reads at, so a 500-token answer takes about 3 seconds.
Of the memory in use, 84% is the weights and 11% is the attention cache. There is no higher-quality file to step up to, so the spare memory is best spent on a longer window.
Reading a 4,000-token prompt, roughly a long email or a few pages of text, takes under a second before the first word appears. That part is limited by compute and not by bandwidth, so it follows a different spec from the writing speed.
The numbers
Llama 3.1 8B Instruct has 8.03 billion parameters across 32 layers. The GeForce RTX 5090 has 32 GB of VRAM at 1792 GB/s, of which about 31.2 GB is left for a model once the operating system has taken its share.
| Quantization | Total needed | Fits | Speed |
|---|---|---|---|
| Q8_0 | 8.83 GB | yes | 157 t/s |
| Q6_K | 7.14 GB | yes | 196 t/s |
| Q5_K_M | 6.38 GB | yes | 221 t/s |
| Q4_K_M | 5.63 GB | yes | 253 t/s |
| IQ4_XS | 5.12 GB | yes | 280 t/s |
| Q3_K_M | 4.82 GB | yes | 299 t/s |
| IQ3_XXS degraded | 4.08 GB | yes | 359 t/s |
| Q2_K degraded | 4.33 GB | yes | 336 t/s |
| IQ2_XXS degraded | 3.21 GB | yes | 471 t/s |
9 of the 9 levels fit entirely. Levels marked degraded are listed for completeness, not as advice.
How long a conversation it holds
Memory rises with the length of the conversation because the attention cache keeps a key and a value for every token. At Q8_0, the longest window that still fits on this device is 128k tokens. Setting the window to what you need and no more is the cheapest way to buy back memory.
| Context | Cache | Total | Fits |
|---|---|---|---|
| 4k | 0.5 GB | 8.33 GB | yes |
| 8k | 1 GB | 8.83 GB | yes |
| 16k | 2 GB | 9.83 GB | yes |
| 32k | 4 GB | 11.83 GB | yes |
| 64k | 8 GB | 15.83 GB | yes |
| 128k | 16 GB | 23.83 GB | yes |
The settings we would use
llama-server -m <model>.gguf -ngl all -c 8192 -fa on
In Ollama, set OLLAMA_CONTEXT_LENGTH to the window you want, run ollama ps while the model is loaded and check that
the processor column reads 100% GPU. If it shows a split on a model that fits, the Ollama guide shows how to fix it.
The llama.cpp flags guide explains each option above.
What else an RTX 5090 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 |
|---|---|---|---|
| Gemma 4 31B Instruct | Q6_K | 27.61 GB | 48.5 t/s |
| Qwen3.5 27B | Q8_0 | 27.32 GB | 49.1 t/s |
| Qwen3 32B | Q6_K | 26.86 GB | 49.9 t/s |
| Qwen3 30B A3B | Q8_0 | 30.99 GB | 190 t/s |
The same model on nearby hardware
Memory decides whether Llama 3.1 8B 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 4090 | 24 GB | 1008 GB/s | Q8_0, 88.0 t/s |
| Apple M3 Max 36GB | 36 GB | 400 GB/s | Q8_0, 32.3 t/s |
See how fast it feels
GeForce RTX 5090 running Llama 3.1 8B Instruct at Q8_0
YouWhy does my model use more memory when the conversation gets longer?
Reading your question
Simulated from our estimate at a 512-token question, not a recording. Assumes nothing else is competing for the GPU.
Questions people ask
How much VRAM does Llama 3.1 8B Instruct need on RTX 5090?
At Q4_K_M with an 8k context window it needs 5.63 GB in total: 4.22 GB of weights, 1 GB of KV cache and 0.41 GB of runtime overhead. The GeForce RTX 5090 makes about 31.2 GB available once the operating system has taken its share.
What context length can RTX 5090 handle with Llama 3.1 8B Instruct?
At Q8_0 the longest window that still fits is about 128k tokens. Memory grows with the window, so a shorter one leaves room for a higher quantization.
How fast is Llama 3.1 8B Instruct on RTX 5090?
We estimate about 157 tokens per second at Q8_0, which is faster than most people read. A 500-token answer takes about 3 seconds. Memory bandwidth of 1792 GB/s sets that ceiling, and the estimate is good to about 10 to 30 percent.
Keep going
Everything the GeForce RTX 5090 runs, the full breakdown for Llama 3.1 8B Instruct, how the numbers are made in the methodology, or a different pairing in the calculator.