RunsOnMy ModelsGPUs & MacsBy VRAMAPI pricesVRAM calculatorNew

Can SmolVLM Instruct run on the DGX Spark?

Yes. At 8K context the best fit is F16, using about 5.4 GB of the 109 GB usable on the DGX Spark. Estimated speed ceiling: ~75 tokens/sec.

Memory by quant and context

Quant4K ctx8K ctxMax tok/s (est., 8K)
2-bit (Q2_K)2.0 GB Fits2.8 GB Fits324
3-bit (Q3_K_M)2.3 GB Fits3.0 GB Fits249
4-bit (Q4_K_M) •2.3 GB Fits3.0 GB Fits245
5-bit (Q5_K_M)2.7 GB Fits3.5 GB Fits171
6-bit (Q6_K)3.0 GB Fits3.7 GB Fits148
8-bit (Q8_0) •3.0 GB Fits3.8 GB Fits142
16-bit (F16/BF16) •4.6 GB Fits5.4 GB Fits75

• real GGUF file size from ggml-org/SmolVLM-Instruct-GGUF. Speed is a bandwidth ceiling (273 GB/s), not a benchmark. Methodology.

More

SmolVLM Instruct on every GPU and Mac · Every model that runs on the DGX Spark · Other models that fit 64 GB VRAM · Try other settings in the calculator

Data: Hugging Face Hub (HuggingFaceTB/SmolVLM-Instruct), NVIDIA specs. Data as of 2026-11-17.