GPUFits
Can RTX 5090 run Llama 3.1 8B?
✅ Comfortable
Llama 3.1 8B @ Q4_K_M · 8K context
Needed
7.5 GB
Usable
32.0 GB
VRAM breakdown
| Weights (Q4_K_M) | 4.9 GB |
| KV cache (8K context) | 1.1 GB |
| Runtime overhead | 1.5 GB |
| Total | 7.5 GB |
Computed at 8K context with fp16 KV cache. Longer contexts need more VRAM — use the VRAM calculator for other settings.
Recommended quantization
FP16
Estimated generation speed
~84 tok/s @ FP16
Theoretical estimates based on published architecture data and measured GGUF sizes; real-world speed varies ±30%.
Cheaper GPUs that run it
Smaller models this GPU runs well
FAQ
- How much VRAM does Llama 3.1 8B need?
- At Q4_K_M with 8K context: 4.9GB weights + 1.1GB KV cache + 1.5GB runtime overhead = 7.5GB total. RTX 5090 offers 32.0GB usable VRAM, so the verdict is: Comfortable.
- What is the best quantization for Llama 3.1 8B on RTX 5090?
- FP16 — the highest tier that still fits within 32.0GB usable VRAM at 8K context. Lower tiers (Q3/Q2) fit too but cost noticeable quality.
- How fast does Llama 3.1 8B run on RTX 5090?
- About 84 tokens/s at FP16 (theoretical estimate, ±30% in real-world use).
Check another combination in the GPU Checker →
Data verified 2026-08-04