Can I run Llama 3.1 8B on an RTX 5090?
Yes: Llama 3.1 8B needs about 9.88 GB at Q4_K_M with 32K tokens of context, which fits the 32 GB RTX 5090 with 22.1 GB to spare. At 32K the RTX 5090 holds up to BF16 (21.4 GB), and Q4_K_M runs up to 128K (full) tokens.
Yes Q4_K_M with 32K tokens of context
- Q4_K_M, 32K
- 9.88 GB
- RTX 5090
- 32 GB, 1,792 GB/s
- To spare
- 22.1 GB
- Tokens/s
- 93–137 tokens/s
At Q4_K_M with 32K tokens of context it writes about 93–137 tokens/s for one request on an RTX 5090.
Best precision for Llama 3.1 8B on an RTX 5090
The most precise setting that leaves at least 0.5 GB free; one that fits with less is marked tight.
| Context | Best fit | Memory | Free | Tokens/s |
|---|---|---|---|---|
| 8K | BF16 | 18.1 GB | 13.9 GB | 53–75 |
| 32K | BF16 | 21.4 GB | 10.6 GB | 45–64 |
| 128K (full) | Q8_0 | 26.8 GB | 5.16 GB | 36–51 |
Llama 3.1 8B on the RTX 5090 as the context fills
One request, FP16 KV cache, 0.5 GB plus 10% overhead; a minus sign is memory missing, and tight is less than 0.5 GB free.
| Context | Q4_K_M | Free | Tokens/s | Q8_0 | Free | Tokens/s |
|---|---|---|---|---|---|---|
| 4K | 6.03 GB | 26.0 GB | 143–222 | 9.79 GB | 22.2 GB | 93–138 |
| 8K | 6.58 GB | 25.4 GB | 133–204 | 10.3 GB | 21.7 GB | 89–131 |
| 16K | 7.68 GB | 24.3 GB | 116–175 | 11.4 GB | 20.6 GB | 81–118 |
| 32K | 9.88 GB | 22.1 GB | 93–137 | 13.6 GB | 18.4 GB | 69–100 |
| 64K | 14.3 GB | 17.7 GB | 66–95 | 18.0 GB | 14.0 GB | 53–76 |
| 128K (full) | 23.1 GB | 8.92 GB | 42–59 | 26.8 GB | 5.16 GB | 36–51 |
Llama 3.1 8B on more than one RTX 5090
| Cards | Best at 32K | Tokens/s | Q4_K_M longest | Q4_K_M tokens/s | Rent per hour |
|---|---|---|---|---|---|
| 2× (64 GB) | BF16 | 72–113 | 128K (full) | 123–213 | $1.38 |
| 4× (128 GB) | BF16 | 115–197 | 128K (full) | 172–333 | $2.76 |
Tensor parallel with the combined bandwidth, as in vLLM; llama.cpp splits layers by default and runs at about one card's speed.
Other options
- The next smaller setting, Q3_K_M, takes 8.92 GB at 32K, 23.1 GB under the RTX 5090; it fits with 0.5 GB to spare up to 128K (full) tokens.
- Renting an RTX 5090 costs about $0.69 an hour (median on getdeploying.com, 2026-09-29): $1.4–2.1 per million tokens at 93–137 tokens/s.
- Smallest setup for Q4_K_M at 32K: RTX 3060 12GB.
Run Llama 3.1 8B on the RTX 5090 with llama-server
llama-server -hf bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M -c 131072 -ngl 99Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf, 4.9 GB, from bartowski/
Questions
Can I run Llama 3.1 8B on an RTX 5090?
Yes: Llama 3.1 8B needs about 9.88 GB at Q4_K_M with 32K tokens of context, which fits the 32 GB RTX 5090 with 22.1 GB to spare. At 32K the RTX 5090 holds up to BF16 (21.4 GB), and Q4_K_M runs up to 128K (full) tokens.
How fast is Llama 3.1 8B on an RTX 5090?
At Q4_K_M with 32K tokens of context it writes about 93–137 tokens/s for one request on an RTX 5090.
What does a second RTX 5090 change for Llama 3.1 8B?
Two RTX 5090 cards (64 GB in one tensor-parallel group) hold Llama 3.1 8B at BF16 with 32K, and Q4_K_M up to 128K (full) tokens, at about 123–213 tokens/s.
Try other settings in the VRAM calculator or the speed calculator. See also Llama 3.1 8B VRAM requirements, what LLMs an RTX 5090 can run and every pair, or detect your own GPU. Model data checked .