GeForce RTX 5090 · Llama 3.3 70B Instruct · Q4_K_M
Only with CPU offload. Llama 3.3 70B Instruct at Q4_K_M does not fit in the 32 GB on GeForce RTX 5090, and a 32 GB system RAM pool runs it slowly, roughly 2–3 tokens per second.
| Piece | Value | Note |
|---|---|---|
| Weights | 39.6 GiB | Llama-3.3-70B-Instruct-Q4_K_M.gguf |
| KV cache | 2.50 GiB | 8,192 tokens |
| Engine overhead | 0.50 GiB | Flat 0.5 GiB allowance |
| 5% headroom | 2.13 GiB | On the sum above |
| Total | 44.7 GiB | Offload |
| Fast memory | 32.0 GiB | 32 GB VRAM |
| Speed | 2–3 tok/s | Dense bandwidth range |
ollama run llama3.3:70b
GeForce RTX 5090 · Llama 3.3 70B Instruct ·GGUF file · Methodology
FAQ
How was this fit calculated?
Weights 39.6 GiB, KV cache 2.50 GiB at 8,192 tokens, engine overhead 0.50 GiB, then 5% headroom, for 44.7 GiB. Fast memory available is 32.0 GiB.
What if I change the quant or the context?
This page is Q4_K_M at 8,192 tokens. A smaller quant or a shorter context lowers the total. The homepage calculator changes both without a new page.
Can I trust the tokens per second?
The range 2–3 tok/s is a bandwidth roofline, 55% to 80% efficient. This dense model reads the whole weight file each token. Measure your own box before you buy around it.
What do I type?
ollama run llama3.3:70b
Is GeForce RTX 5090 assumed to have extra RAM?
Yes. Discrete GPU pages assume 32 GB of system RAM. Offload uses whatever remains after 8 GB for Windows, and it is limited by the slower of DDR5-5600 and the PCIe link.