Can Llama 3.2 1B run on the H100 SXM?
Yes. At 8K context the best fit is F16, using about 3.2 GB of the 76.0 GB usable on the H100 SXM. Estimated speed ceiling: ~1355 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) • | 1.4 GB Fits | 1.5 GB Fits | 2.3 GB Fits | 5.3 GB Fits | 5022 |
| 3-bit (Q3_K_M) • | 1.5 GB Fits | 1.6 GB Fits | 2.4 GB Fits | 5.4 GB Fits | 4168 |
| 4-bit (Q4_K_M) • | 1.7 GB Fits | 1.8 GB Fits | 2.5 GB Fits | 5.5 GB Fits | 3506 |
| 5-bit (Q5_K_M) • | 1.8 GB Fits | 1.9 GB Fits | 2.7 GB Fits | 5.7 GB Fits | 3067 |
| 6-bit (Q6_K) • | 1.9 GB Fits | 2.0 GB Fits | 2.8 GB Fits | 5.8 GB Fits | 2708 |
| 8-bit (Q8_0) • | 2.3 GB Fits | 2.4 GB Fits | 3.1 GB Fits | 6.1 GB Fits | 2094 |
| 16-bit (F16/BF16) | 3.1 GB Fits | 3.2 GB Fits | 3.9 GB Fits | 6.9 GB Fits | 1355 |
• real GGUF file size from QuantFactory/Dolphin3.0-Llama3.2-1B-GGUF. Speed is a bandwidth ceiling (3350 GB/s), not a benchmark. Methodology.
More
Llama 3.2 1B on every GPU and Mac · Every model that runs on the H100 SXM · Other models that fit 64 GB VRAM · Try other settings in the calculator
Data: Hugging Face Hub (meta-llama/Llama-3.2-1B), NVIDIA specs. Data as of 2026-11-17.