Can I run Llama 3.3 70B on the GeForce RTX 3090?
Won't runQ4_K_M fits in memory (45.2 GB) but is too slow to use: est. 1.8 tok/s (1.4–2.1, calibrated estimate ±20%). IQ4_XS works instead: Runs slowly, est. 2.2 tok/s (1.7–2.6, calibrated estimate ±20%).
24 GB VRAM · 936.0 GB/s · FP16 35.6 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. The verdict stays the same, but which file fits in memory depends on how much system RAM you have.
- 16 GBWon't runTry Q2_K (heavy quality loss): Runs slowlyest. 5.3 tok/s (4.3–6.4, calibrated estimate ±20%)
- 32 GBWon't runTry IQ4_XS: Runs slowlyest. 2.2 tok/s (1.7–2.6, calibrated estimate ±20%)(used on this page)
- 64 GBWon't runTry IQ4_XS: Runs slowlyest. 2.2 tok/s (1.7–2.6, calibrated estimate ±20%)
- 96 GBWon't runTry IQ4_XS: Runs slowlyest. 2.2 tok/s (1.7–2.6, calibrated estimate ±20%)
- 128 GBWon't runTry IQ4_XS: Runs slowlyest. 2.2 tok/s (1.7–2.6, calibrated estimate ±20%)
Every quant of Llama 3.3 70B on the GeForce RTX 3090
| Quant | File size | Verdict | Speed | Memory | Runs as | Notes |
|---|---|---|---|---|---|---|
| IQ2_M | 24.12 GB | Runs slowly | est. 7.5 tok/s6.0–9.0calibrated estimate ±20% | 24.0 / 24.0 GB + 4.0 GB RAM | Partial offload | — |
| Q2_K | 26.38 GB | Runs slowly | est. 5.3 tok/s4.3–6.4calibrated estimate ±20% | 24.0 / 24.0 GB + 6.2 GB RAM | Partial offload | — |
| IQ4_XS | 37.90 GB | Runs slowly | est. 2.2 tok/s1.7–2.6calibrated estimate ±20% | 24.0 / 24.0 GB + 17.8 GB RAM | Partial offload | — |
| Q4_K_Mbaseline | 42.52 GB | Won't run | est. 1.8 tok/s1.4–2.1calibrated estimate ±20% | 24.0 / 24.0 GB + 22.4 GB RAM | Partial offload | Fits, but too slow to use |
| Q8_0 | 74.98 GB | Won't run | — | Needs about 77.1 GB; 23.4 GB of VRAM and 28.0 GB of RAM are free | — | — |
Where the memory goes at IQ4_XS
- Weights
- 20.1 GB
- KV cache
- 2.7 GB
- Compute buffer
- 0.6 GB
- OS reserve
- 0.6 GB
- Weights in system RAM
- 17.8 GB
System RAM, context length and KV cache
| Context | KV f16 | KV q8_0 | KV q4_0 |
|---|---|---|---|
| 4k | 39.2 GBRuns slowly est. 2.4 tok/s1.9–2.9calibrated estimate ±20% | 38.6 GBRuns slowly est. 2.5 tok/s2.0–3.1calibrated estimate ±20% | 38.3 GBRuns slowly est. 2.6 tok/s2.1–3.1calibrated estimate ±20% |
| 8k | 40.6 GBRuns slowly est. 2.2 tok/s1.7–2.6calibrated estimate ±20% | 39.3 GBRuns slowly est. 2.4 tok/s1.9–2.9calibrated estimate ±20% | 38.7 GBRuns slowly est. 2.5 tok/s2.0–3.0calibrated estimate ±20% |
| 16k | 43.3 GBWon't run est. 1.8 tok/s1.4–2.2calibrated estimate ±20% | 40.9 GBRuns slowly est. 2.1 tok/s1.7–2.6calibrated estimate ±20% | 39.5 GBRuns slowly est. 2.4 tok/s1.9–2.9calibrated estimate ±20% |
| 32k | 48.9 GBWon't run est. 1.3 tok/s1.0–1.6calibrated estimate ±20% | 43.9 GBWon't run est. 1.8 tok/s1.4–2.1calibrated estimate ±20% | 41.2 GBRuns slowly est. 2.1 tok/s1.7–2.5calibrated estimate ±20% |
| 64k | 59.9 GBdoes not fit | 49.9 GBWon't run est. 1.2 tok/s1.0–1.5calibrated estimate ±20% | 44.6 GBWon't run est. 1.7 tok/s1.4–2.0calibrated estimate ±20% |
| 128k | 82.0 GBdoes not fit | 62.0 GBdoes not fit | 51.3 GBdoes not fit |
S = Runs great · A = Runs well · B = Runs slowly · F = Won't run
At IQ4_XS the verdict stays Runs slowly up to 8k context with an f16 KV cache, and up to 16k with q8_0.
Estimated speed
- Generation
- est. 2.2 tok/s (1.7–2.6, 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 3090
Runs wellK2-Horizon MoVA 36B-A4B: Runs well, est. 79.2 tok/s (55.5–103.0, theoretical estimate ±30%)
A lower quant of Llama 3.3 70B
Runs slowlyIQ4_XS (37.90 GB): Runs slowly, est. 2.2 tok/s (1.7–2.6, calibrated estimate ±20%)
How to run it
Commands for the IQ4_XS 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-IQ4_XS.gguf (37.90 GB) from bartowski/Llama-3.3-70B-Instruct-GGUF.
llama-server -hf bartowski/Llama-3.3-70B-Instruct-GGUF:IQ4_XS -c 8192 -ngl 42 -fa on- -ngl 42 puts 42 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:IQ4_XS- 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 3090
Llama 3.3 70B on other hardware
Frequently asked questions
Can I run Llama 3.3 70B on the GeForce RTX 3090?
Q4_K_M fits in memory (45.2 GB) but is too slow to use: est. 1.8 tok/s (1.4–2.1, calibrated estimate ±20%). IQ4_XS works instead: Runs slowly, est. 2.2 tok/s (1.7–2.6, calibrated estimate ±20%). IQ4_XS needs 40.6 GB at 8k context; this setup has 23.4 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 3090?
At IQ4_XS the verdict stays Runs slowly up to 8k context with an f16 KV cache (2.7 GB of KV) and up to 16k with a q8_0 KV cache (2.9 GB). The model supports up to 128k.
Which quant should I use, and how fast is it?
IQ4_XS (37.90 GB) is the recommended balance: Runs slowly, est. 2.2 tok/s (1.7–2.6, 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.