Can Llama 3.2 3B Instruct run on the RTX 6000 Ada?
Yes. At 8K context the best fit is F16, using about 13.5 GB of the 45.6 GB usable on the RTX 6000 Ada. 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) • | 2.4 GB Fits | 2.8 GB Fits | 5.4 GB Fits | 15.9 GB Fits | 704 |
| 3-bit (Q3_K_M) • | 2.7 GB Fits | 3.1 GB Fits | 5.7 GB Fits | 16.2 GB Fits | 569 |
| 4-bit (Q4_K_M) • | 3.0 GB Fits | 3.4 GB Fits | 6.0 GB Fits | 16.5 GB Fits | 475 |
| 5-bit (Q5_K_M) • | 3.2 GB Fits | 3.7 GB Fits | 6.3 GB Fits | 16.8 GB Fits | 413 |
| 6-bit (Q6_K) • | 3.5 GB Fits | 4.0 GB Fits | 6.6 GB Fits | 17.1 GB Fits | 363 |
| 8-bit (Q8_0) • | 4.3 GB Fits | 4.7 GB Fits | 7.3 GB Fits | 17.8 GB Fits | 281 |
| 16-bit (F16/BF16) • | 13.1 GB Fits | 13.5 GB Fits | 16.1 GB Fits | 26.6 GB Fits | 75 |
• real GGUF file size from unsloth/Llama-3.2-3B-Instruct-GGUF. Speed is a bandwidth ceiling (960 GB/s), not a benchmark. Methodology.
More
Llama 3.2 3B Instruct on every GPU and Mac · Every model that runs on the RTX 6000 Ada · Other models that fit 48 GB VRAM · Try other settings in the calculator
Data: Hugging Face Hub (meta-llama/Llama-3.2-3B-Instruct), NVIDIA specs. Data as of 2026-11-17.