Can gemma 3 27b it run on the H200?
Yes. At 8K context the best fit is F16, using about 52.2 GB of the 134 GB usable on the H200. Estimated speed ceiling: ~89 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) • | 11.4 GB Fits | 11.7 GB Fits | 13.6 GB Fits | 21.1 GB Fits | 457 |
| 3-bit (Q3_K_M) • | 14.1 GB Fits | 14.4 GB Fits | 16.3 GB Fits | 23.8 GB Fits | 357 |
| 4-bit (Q4_K_M) • | 17.0 GB Fits | 17.3 GB Fits | 19.2 GB Fits | 26.7 GB Fits | 290 |
| 5-bit (Q5_K_M) • | 19.6 GB Fits | 19.9 GB Fits | 21.8 GB Fits | 29.3 GB Fits | 249 |
| 6-bit (Q6_K) • | 22.3 GB Fits | 22.6 GB Fits | 24.5 GB Fits | 32.0 GB Fits | 217 |
| 8-bit (Q8_0) • | 28.4 GB Fits | 28.7 GB Fits | 30.5 GB Fits | 38.0 GB Fits | 167 |
| 16-bit (F16/BF16) • | 51.9 GB Fits | 52.2 GB Fits | 54.1 GB Fits | 61.6 GB Fits | 89 |
• real GGUF file size from unsloth/gemma-3-27b-it-GGUF. Speed is a bandwidth ceiling (4800 GB/s), not a benchmark. Methodology.
More
gemma 3 27b it on every GPU and Mac · Every model that runs on the H200 · Other models that fit 64 GB VRAM · Try other settings in the calculator
Data: Hugging Face Hub (google/gemma-3-27b-it), NVIDIA specs. Data as of 2026-11-17.