InternVL3 5 30B A3B is a mixture-of-experts vision-language model that activates only 3 billion parameters per forward pass despite housing 30 billion total, making it punchy and efficient at inference time. It handles both text and image inputs fluidly, parsing diagrams, charts, and natural scenes alongside written prompts. The sparse activation pattern means it can tackle visually rich tasks without the compute overhead you'd expect from a model its full size.