Can deepseek moe 16b base run on the H200?
Yes. At 8K context the best fit is F16, using about 32.8 GB of the 134 GB usable on the H200. Estimated speed ceiling: ~848 tokens/sec.
Memory by quant and context
| Quant | 4K ctx | Max tok/s (est., 8K) |
|---|---|---|
| 2-bit (Q2_K) • | 7.7 GB Fits | 4143 |
| 3-bit (Q3_K_M) • | 9.4 GB Fits | 3277 |
| 4-bit (Q4_K_M) • | 11.6 GB Fits | 2561 |
| 5-bit (Q5_K_M) • | 13.0 GB Fits | 2243 |
| 6-bit (Q6_K) • | 15.1 GB Fits | 1895 |
| 8-bit (Q8_0) • | 17.7 GB Fits | 1596 |
| 16-bit (F16/BF16) | 32.0 GB Fits | 848 |
• real GGUF file size from mradermacher/deepseek-moe-16b-base-GGUF. Speed is a bandwidth ceiling (4800 GB/s), not a benchmark. Methodology.
More
deepseek moe 16b base 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 (deepseek-ai/deepseek-moe-16b-base), NVIDIA specs. Data as of 2026-11-17.