Can Kimi Dev 72B run on the H200?
Yes. At 8K context the best fit is Q8_0, using about 75.1 GB of the 134 GB usable on the H200. Estimated speed ceiling: ~62 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) | 27.3 GB Fits | 28.6 GB Fits | 36.1 GB Fits | 66.1 GB Fits | 176 |
| 3-bit (Q3_K_M) • | 37.0 GB Fits | 38.3 GB Fits | 45.8 GB Fits | 75.8 GB Fits | 127 |
| 4-bit (Q4_K_M) | 43.3 GB Fits | 44.6 GB Fits | 52.1 GB Fits | 82.1 GB Fits | 108 |
| 5-bit (Q5_K_M) | 50.2 GB Fits | 51.4 GB Fits | 58.9 GB Fits | 88.9 GB Fits | 93 |
| 6-bit (Q6_K) • | 61.9 GB Fits | 63.1 GB Fits | 70.6 GB Fits | 101 GB Fits | 75 |
| 8-bit (Q8_0) • | 73.9 GB Fits | 75.1 GB Fits | 82.6 GB Fits | 113 GB Fits | 62 |
| 16-bit (F16/BF16) • | 137 GB No | 139 GB No | 146 GB No | 176 GB No | — |
• real GGUF file size from unsloth/Kimi-Dev-72B-GGUF. Speed is a bandwidth ceiling (4800 GB/s), not a benchmark. Methodology.
More
Kimi Dev 72B 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 (moonshotai/Kimi-Dev-72B), NVIDIA specs. Data as of 2026-11-17.