Llama 3.3 70B Instruct at Q4_K_M needs about 45.1 GB, more than the 32 GB of the RTX 5090: part of the model runs from system RAM, which slows generation. At Q3_K_M it does not fit in VRAM.
Q4_K_M · 8K context
The model needs about 45.1 GB but only 31 GB of VRAM is usable: 14 GB must be loaded into system RAM. It runs, 69% on the GPU, but more slowly.
Estimated speed is at most 2.5 tokens/s: usable, but slow.
~45.1 GB (43.9 GB–46.3 GB) needed · estimate
14 GB must be offloaded to system RAM
EstimateData updated 24 days agoSource: Run LLM Locally engine, Hugging Face config.json + model card
What makes this estimate less reliable
| Llama 3.3 70B Instruct Q8_08K~76.8 GB · 0.4–0.8 tok/s estimate | Not recommended |
|---|---|
| Llama 3.3 70B Instruct Q6_K8K~60.1 GB · 0.7–1.3 tok/s estimate | Not recommended |
| Llama 3.3 70B Instruct Q5_K_M8K~52.3 GB · 0.9–1.8 tok/s estimate | Not recommended |
| Llama 3.3 70B Instruct Q4_K_M8K~45.1 GB · 1.4–2.5 tok/s estimate | Possible but slow |
| Llama 3.3 70B Instruct IQ4_XS8K~40.6 GB · 2–4 tok/s estimate | Possible but slow |
| Llama 3.3 70B Instruct Q3_K_M8K~37 GB · 3–6 tok/s estimate | CPU/GPU offload |
Prices and links follow the market you choose.
NVIDIA
CompatibleRuns Llama 3.3 70B Instruct Q4_K_M entirely in memory
AMD
CompatibleRuns Llama 3.3 70B Instruct Q4_K_M entirely in memory
Unified memory PC
CompatibleRuns Llama 3.3 70B Instruct Q4_K_M entirely in memory
No current offer imported for this market: the link opens a marketplace search.
