Can deepseek vl2 tiny run on the H100 PCIe?
Yes. At 8K context the best fit is F16, using about 7.4 GB of the 76.0 GB usable on the H100 PCIe. Estimated speed ceiling: ~851 tokens/sec.
Memory by quant and context
| Quant | 4K ctx | Max tok/s (est., 8K) |
|---|---|---|
| 2-bit (Q2_K) | 2.1 GB Fits | 4538 |
| 3-bit (Q3_K_M) | 2.4 GB Fits | 3482 |
| 4-bit (Q4_K_M) | 2.8 GB Fits | 2784 |
| 5-bit (Q5_K_M) | 3.1 GB Fits | 2388 |
| 6-bit (Q6_K) | 3.5 GB Fits | 2075 |
| 8-bit (Q8_0) | 4.2 GB Fits | 1601 |
| 16-bit (F16/BF16) | 7.2 GB Fits | 851 |
• real GGUF file size. Speed is a bandwidth ceiling (2000 GB/s), not a benchmark. Methodology.
More
deepseek vl2 tiny on every GPU and Mac · Every model that runs on the H100 PCIe · Other models that fit 64 GB VRAM · Try other settings in the calculator
Data: Hugging Face Hub (deepseek-ai/deepseek-vl2-tiny), NVIDIA specs. Data as of 2026-11-17.