RunsOnMy ModelsGPUs & MacsBy VRAMAPI pricesVRAM calculatorNew

Can WeVisOne VL 4B run on the H100 PCIe?

Yes. At 8K context the best fit is F16, using about 10.0 GB of the 76.0 GB usable on the H100 PCIe. Estimated speed ceiling: ~226 tokens/sec.

Memory by quant and context

Quant4K ctx8K ctx32K ctx128K ctxMax tok/s (est., 8K)
2-bit (Q2_K) •2.9 GB Fits3.5 GB Fits6.9 GB Fits20.4 GB Fits1112
3-bit (Q3_K_M) •3.3 GB Fits3.9 GB Fits7.3 GB Fits20.8 GB Fits891
4-bit (Q4_K_M) •3.8 GB Fits4.3 GB Fits7.7 GB Fits21.2 GB Fits736
5-bit (Q5_K_M) •4.2 GB Fits4.8 GB Fits8.1 GB Fits21.6 GB Fits633
6-bit (Q6_K) •4.6 GB Fits5.2 GB Fits8.6 GB Fits22.1 GB Fits551
8-bit (Q8_0) •5.6 GB Fits6.2 GB Fits9.6 GB Fits23.1 GB Fits426
16-bit (F16/BF16) •9.5 GB Fits10.0 GB Fits13.4 GB Fits26.9 GB Fits226

• real GGUF file size from mradermacher/WeVisOne-VL-4B-GGUF. Speed is a bandwidth ceiling (2000 GB/s), not a benchmark. Methodology.

More

WeVisOne VL 4B 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 (tencent/WeVisOne-VL-4B), NVIDIA specs. Data as of 2026-11-17.