Yes, I would say that it's closer to your first approach (attributes in a key/value table) than to your second (attributes in a blob you have to decode yourself).

Another couple of reasons that completely separate tables might be better: there is no redundant attribute-key field to store, and you avoid attribute-key spelling errors (like "color" vs. "colour"). If the attributes are entered through a user interface, how are you going to check for valid attribute names?