Can Llama Guard 4 12B run on the RTX 5090?
Yes. At 8K context the best fit is F16, using about 24.6 GB of the 30.4 GB usable on the RTX 5090. Estimated speed ceiling: ~75 tokens/sec.
Memory by quant and context
| Quant | 4K ctx | 8K ctx | 32K ctx | 128K ctx | Max tok/s (est., 8K) |
|---|---|---|---|---|---|
| 2-bit (Q2_K) | 5.7 GB Fits | 6.5 GB Fits | 11.0 GB Fits | 29.0 GB Tight | 398 |
| 3-bit (Q3_K_M) | 7.0 GB Fits | 7.7 GB Fits | 12.2 GB Fits | 30.2 GB Tight | 306 |
| 4-bit (Q4_K_M) | 8.4 GB Fits | 9.1 GB Fits | 13.6 GB Fits | 31.6 GB No | 244 |
| 5-bit (Q5_K_M) | 9.5 GB Fits | 10.2 GB Fits | 14.7 GB Fits | 32.7 GB No | 210 |
| 6-bit (Q6_K) | 10.7 GB Fits | 11.5 GB Fits | 16.0 GB Fits | 34.0 GB No | 182 |
| 8-bit (Q8_0) | 13.4 GB Fits | 14.2 GB Fits | 18.7 GB Fits | 36.7 GB No | 141 |
| 16-bit (F16/BF16) | 23.9 GB Fits | 24.6 GB Fits | 29.1 GB Tight | 47.1 GB No | 75 |
• real GGUF file size. Speed is a bandwidth ceiling (1792 GB/s), not a benchmark. Methodology.
More
Llama Guard 4 12B on every GPU and Mac · Every model that runs on the RTX 5090 · Other models that fit 32 GB VRAM · Try other settings in the calculator
Data: Hugging Face Hub (meta-llama/Llama-Guard-4-12B), NVIDIA specs. Data as of 2026-11-17.