Can I Run Llama 3.3 on NVIDIA GeForce RTX 5090?

Name: LLM Configurator — GPU VRAM Checker
Author: LLM Configurator

Superseded model. Llama 3.3 has been superseded by Llama 4. This page is kept for reference; the newer family is a better starting point. View Llama 4 →

作者： Jakub Rusinowski · 最后更新： 2026年7月12日

Yes, comfortably — you'll have ~11.8 GB of headroom running Llama 3.3 70B Instruct at Q2_K_XS (Tight) (20.2 GB, ~61 tok/s (est.)).

See what else this hardware can run →

购买此硬件 NVIDIA GeForce RTX 5090 32GB — 32 GB VRAM · 575 W board power 立即云端部署 RunPod 上的 RTX 5090

或在 Vast.ai 比较

作为亚马逊联盟成员，我们从符合条件的购买中获得收入。云 GPU 链接为推荐链接——我们可能获得佣金，您无需额外付费。

联盟营销声明: 本页部分链接为联盟推广链接——如果你通过它们购买，LLM Configurator 可能会获得佣金，而你无需支付任何额外费用。作为亚马逊联盟成员（Amazon Associate），LLM Configurator 会从符合条件的购买中获得收益。

NVIDIA GeForce RTX 5090 32GB

32 GB VRAM · 575 W board power

2026年价格波动较大——请以当前商品页价格为准。

在亚马逊查看价格

NVIDIA GeForce RTX 5090 Specs

VRAM	32 GB
Memory Bandwidth	1792 GB/s

Llama 3.3 70B Instruct on the NVIDIA GeForce RTX 5090: VRAM by quantization

Quant	VRAM needed	Fits 32 GB?	Max context	Whole PC
F16	142.1 GB	✗ No	—	—
Q8_0	76.5 GB	✗ No	—	—
Q6_K	59.5 GB	✗ No	—	—
Q5_K_M	51.8 GB	✗ No	—	—
Q4_K_M	44.4 GB	✗ No	—	—
Q3_K_M	32 GB	✓ Yes	4K	—
Q2_K	25.2 GB	✓ Yes	16K	—

Assumes a 4K-token context with an f16 KV cache. A longer window needs more; a quantized KV cache needs less. “Max context” is the largest window that still fits in 32 GB. Figures are estimates from parameter count, quantization and memory bandwidth — the analyzer lets you tune KV-cache quant and context.

Context costs VRAM too. At Q2_K_XS (Tight) on the NVIDIA GeForce RTX 5090: 8K 23.7 GB · 32K 31.7 GB · 128K 64 GB ✗. Past 128K it no longer fits 32 GB — a q8_0 KV cache buys roughly half of that back, which the calculator will price for you.

No compatible GPU? Llama 3.3 70B Instruct on 256 GB of system RAM, CPU only: runs, but too slowly to be worth it.