Vision engine
Can I run this vision model?
The vision encoder adds resident weights, and images add context tokens and KV cache. See the estimated memory for your model, hardware and image inputs.
Model and hardware
Estimate resident memory for a decoder-only vision-language model.
Qwen · Apache 2.0
Hardware picker is loading. Please wait.
Context and precision
The vision encoder stays in fp16; quantization applies to the text weights.
Pick your hardware above to see if Qwen3-VL 2B fits.
Images become context
The vision encoder adds resident weights. Each image becomes hundreds to thousands of tokens in the context, which use KV cache just like a long text prompt. The breakdown separates text weights, the encoder, text KV and image KV so the source of the memory cost stays visible.
Text-only model? Back to the main calculator →