Can Qwen2.5 VL 7B Instruct run on the H200?
Yes. At 8K context the best fit is F16, using about 15.3 GB of the 134 GB usable on the H200. Estimated speed ceiling: ~315 tokens/sec.
Memory by quant and context
| Quant | 4K ctx | 8K ctx | 32K ctx | Max tok/s (est., 8K) |
|---|---|---|---|---|
| 2-bit (Q2_K) • | 3.7 GB Fits | 3.9 GB Fits | 5.2 GB Fits | 1592 |
| 3-bit (Q3_K_M) • | 4.5 GB Fits | 4.7 GB Fits | 6.0 GB Fits | 1260 |
| 4-bit (Q4_K_M) • | 5.3 GB Fits | 5.5 GB Fits | 6.8 GB Fits | 1025 |
| 5-bit (Q5_K_M) • | 6.0 GB Fits | 6.2 GB Fits | 7.5 GB Fits | 882 |
| 6-bit (Q6_K) • | 6.7 GB Fits | 7.0 GB Fits | 8.3 GB Fits | 767 |
| 8-bit (Q8_0) • | 8.5 GB Fits | 8.7 GB Fits | 10.0 GB Fits | 593 |
| 16-bit (F16/BF16) • | 15.1 GB Fits | 15.3 GB Fits | 16.6 GB Fits | 315 |
• real GGUF file size from unsloth/Qwen2.5-VL-7B-Instruct-GGUF. Speed is a bandwidth ceiling (4800 GB/s), not a benchmark. Methodology.
More
Qwen2.5 VL 7B Instruct 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 (Qwen/Qwen2.5-VL-7B-Instruct), NVIDIA specs. Data as of 2026-11-17.