Can Llama 3.2 11B Vision Instruct run on the 2x RTX 5090?
Yes. At 8K context the best fit is F16, using about 20.1 GB of the 60.8 GB usable on the 2x RTX 5090. Estimated speed ceiling: ~92 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) | 5.0 GB Fits | 5.6 GB Fits | 9.4 GB Fits | 24.4 GB Fits | 448 |
| 3-bit (Q3_K_M) | 6.1 GB Fits | 6.8 GB Fits | 10.5 GB Fits | 25.5 GB Fits | 344 |
| 4-bit (Q4_K_M) • | 6.8 GB Fits | 7.4 GB Fits | 11.2 GB Fits | 26.2 GB Fits | 301 |
| 5-bit (Q5_K_M) | 8.4 GB Fits | 9.0 GB Fits | 12.7 GB Fits | 27.7 GB Fits | 236 |
| 6-bit (Q6_K) | 9.4 GB Fits | 10.0 GB Fits | 13.8 GB Fits | 28.8 GB Fits | 205 |
| 8-bit (Q8_0) • | 11.0 GB Fits | 11.6 GB Fits | 15.3 GB Fits | 30.3 GB Fits | 172 |
| 16-bit (F16/BF16) • | 19.5 GB Fits | 20.1 GB Fits | 23.9 GB Fits | 38.9 GB Fits | 92 |
• real GGUF file size from leafspark/Llama-3.2-11B-Vision-Instruct-GGUF. Speed is a bandwidth ceiling (1792 GB/s), not a benchmark. Methodology.
More
Llama 3.2 11B Vision Instruct on every GPU and Mac · Every model that runs on the 2x RTX 5090 · Try other settings in the calculator
Data: Hugging Face Hub (meta-llama/Llama-3.2-11B-Vision-Instruct), NVIDIA specs. Data as of 2026-11-17.