Can I run Llama 3.3 70B on the GeForce RTX 5090?
Runs slowlyQ4_K_M does not fit fully in VRAM, so 14.4 GB of the weights run from system RAM: Runs slowly, est. 2.7 tok/s (2.2–3.3, calibrated estimate ±20%). Q2_K (heavy quality loss) would reach Runs great, est. 43.2 tok/s (38.0–48.3, calibrated estimate ±12%).
32 GB VRAM · 1,792.0 GB/s · FP16 104.8 TFLOPS. Assumes 32 GB of DDR5-5600 system RAM, Windows with this GPU driving the display, 8k context and an f16 KV cache.
Verdict by system RAM
This page assumes 32 GB of system RAM. For this pair, the verdict depends on how much system RAM you have.
- 16 GBWon't runTry IQ4_XS: Runs slowlyest. 3.9 tok/s (3.1–4.7, calibrated estimate ±20%)Q2_K (heavy quality loss): Runs great
- 32 GBRuns slowlyest. 2.7 tok/s (2.2–3.3, calibrated estimate ±20%)Q2_K (heavy quality loss): Runs great(used on this page)
- 64 GBRuns slowlyest. 2.7 tok/s (2.2–3.3, calibrated estimate ±20%)Q2_K (heavy quality loss): Runs great
- 96 GBRuns slowlyest. 2.7 tok/s (2.2–3.3, calibrated estimate ±20%)Q2_K (heavy quality loss): Runs great
- 128 GBRuns slowlyest. 2.7 tok/s (2.2–3.3, calibrated estimate ±20%)Q2_K (heavy quality loss): Runs great
Every quant of Llama 3.3 70B on the GeForce RTX 5090
| Quant | File size | Verdict | Speed | Memory | Runs as | Notes |
|---|---|---|---|---|---|---|
| IQ2_M | 24.12 GB | Runs great | est. 46.8 tok/s41.2–52.4calibrated estimate ±12% | 27.5 / 32.0 GB | Full GPU | — |
| Q2_K | 26.38 GB | Runs great | est. 43.2 tok/s38.0–48.3calibrated estimate ±12% | 29.9 / 32.0 GB | Full GPU | — |
| IQ4_XS | 37.90 GB | Runs slowly | est. 3.9 tok/s3.1–4.7calibrated estimate ±20% | 32.0 / 32.0 GB + 9.8 GB RAM | Partial offload | — |
| Q4_K_Mbaseline | 42.52 GB | Runs slowly | est. 2.7 tok/s2.2–3.3calibrated estimate ±20% | 32.0 / 32.0 GB + 14.4 GB RAM | Partial offload | — |
| Q8_0 | 74.98 GB | Won't run | — | Needs about 77.1 GB; 31.4 GB of VRAM and 28.0 GB of RAM are free | — | — |
Where the memory goes at Q4_K_M
- Weights
- 28.1 GB
- KV cache
- 2.7 GB
- Compute buffer
- 0.6 GB
- OS reserve
- 0.6 GB
- Weights in system RAM
- 14.4 GB
System RAM, context length and KV cache
| Context | KV f16 | KV q8_0 | KV q4_0 |
|---|---|---|---|
| 4k | 43.8 GBRuns slowly est. 3.1 tok/s2.5–3.7calibrated estimate ±20% | 43.2 GBRuns slowly est. 3.3 tok/s2.6–3.9calibrated estimate ±20% | 42.8 GBRuns slowly est. 3.4 tok/s2.7–4.1calibrated estimate ±20% |
| 8k | 45.2 GBRuns slowly est. 2.7 tok/s2.2–3.3calibrated estimate ±20% | 43.9 GBRuns slowly est. 3.1 tok/s2.4–3.7calibrated estimate ±20% | 43.3 GBRuns slowly est. 3.3 tok/s2.6–3.9calibrated estimate ±20% |
| 16k | 47.9 GBRuns slowly est. 2.2 tok/s1.8–2.6calibrated estimate ±20% | 45.4 GBRuns slowly est. 2.7 tok/s2.1–3.2calibrated estimate ±20% | 44.1 GBRuns slowly est. 3.0 tok/s2.4–3.6calibrated estimate ±20% |
| 32k | 53.5 GBWon't run est. 1.5 tok/s1.2–1.8calibrated estimate ±20% | 48.5 GBRuns slowly est. 2.1 tok/s1.7–2.6calibrated estimate ±20% | 45.8 GBRuns slowly est. 2.6 tok/s2.1–3.1calibrated estimate ±20% |
| 64k | 64.5 GBdoes not fit | 54.5 GBWon't run est. 1.4 tok/s1.2–1.7calibrated estimate ±20% | 49.1 GBRuns slowly est. 2.0 tok/s1.6–2.4calibrated estimate ±20% |
| 128k | 86.6 GBdoes not fit | 66.6 GBdoes not fit | 55.9 GBWon't run est. 1.4 tok/s1.1–1.6calibrated estimate ±20% |
S = Runs great · A = Runs well · B = Runs slowly · F = Won't run
At Q4_K_M the verdict stays Runs slowly up to 16k context with an f16 KV cache, and up to 32k with q8_0.
Estimated speed
- Generation
- est. 2.7 tok/s (2.2–3.3, calibrated estimate ±20%)
- Prompt processing
- Prompt-processing estimates only apply when the whole model runs on the GPU.
Measured on this exact combination
No public measurement for this exact combination yet; the numbers above are estimates.
If this is not enough
A smaller model that runs well on the GeForce RTX 5090
Runs greatK2-Horizon MoVA 36B-A4B: Runs great, est. 149.6 tok/s (104.7–194.5, theoretical estimate ±30%)
A lower quant of Llama 3.3 70B
Runs greatQ2_K (26.38 GB, heavy quality loss): Runs great, est. 43.2 tok/s (38.0–48.3, calibrated estimate ±12%)
How to run it
Commands for the Q4_K_M file. Both tools download from Hugging Face on first run.
These commands assume 32 GB of system RAM. For a different amount of RAM, pick it under “System RAM, context length and KV cache” above to see the commands for it.
File: Llama-3.3-70B-Instruct-Q4_K_M.gguf (42.52 GB) from bartowski/Llama-3.3-70B-Instruct-GGUF.
llama-server -hf bartowski/Llama-3.3-70B-Instruct-GGUF:Q4_K_M -c 8192 -ngl 52 -fa on- -ngl 52 puts 52 of 80 layers on the GPU; if VRAM overflows, lower it by 1–2.
- -fa on enables flash attention (needed for KV cache quantization).
- --cache-type-k q8_0 --cache-type-v q8_0 shrinks the KV cache to 1.44 GB at 8k context.
PowerShell (quit the Ollama tray app first):
$env:OLLAMA_CONTEXT_LENGTH="8192"; ollama serve
ollama run hf.co/bartowski/Llama-3.3-70B-Instruct-GGUF:Q4_K_M- Ollama defaults to 4k context on GPUs under 24 GB. Set OLLAMA_CONTEXT_LENGTH when starting the server, or /set parameter num_ctx inside the chat.
- Ollama splits layers between GPU and CPU automatically.
/set parameter num_ctx 8192
Related pages
Guide: How much VRAM do local LLMs need? GPU requirements for 8–32 GB (2026)
Other models on the GeForce RTX 5090
Llama 3.3 70B on other hardware
Frequently asked questions
Can I run Llama 3.3 70B on the GeForce RTX 5090?
Q4_K_M does not fit fully in VRAM, so 14.4 GB of the weights run from system RAM: Runs slowly, est. 2.7 tok/s (2.2–3.3, calibrated estimate ±20%). Q2_K (heavy quality loss) would reach Runs great, est. 43.2 tok/s (38.0–48.3, calibrated estimate ±12%). Q4_K_M needs 45.2 GB at 8k context; this setup has 31.4 GB of usable memory and 28.0 GB of free system RAM.
How much context can Llama 3.3 70B use on the GeForce RTX 5090?
At Q4_K_M the verdict stays Runs slowly up to 16k context with an f16 KV cache (5.4 GB of KV) and up to 32k with a q8_0 KV cache (5.7 GB). The model supports up to 128k.
Which quant should I use, and how fast is it?
Q4_K_M (42.52 GB) is the recommended balance: Runs slowly, est. 2.7 tok/s (2.2–3.3, calibrated estimate ±20%). It is also the largest tracked file that gets this verdict.
Speeds are estimates from memory bandwidth, calibrated against public benchmarks, and each one comes with an error band and a confidence label. Real results vary with drivers, backend, context length and thermals.