Graph-based data are very ubiquitous in many real-world scenarios, and it is usually difficult to mine valuable information from the large scale graphs because of traditional sparse and high-dimension representations of nodes in the graphs. To address this issue, the graph embedding techniques which aim to map the nodes of the graph into a low-dimension dense vector space are proposed. These vectors can act as the features of nodes for many graph analytics tasks. However, most existing graph embedding algorithms highly rely on gradient information, which largely restricts the flexibility and universality of algorithms and easily reaches local optimum. In this paper, we propose a general and flexible gradient-free (e.g. particle swarm optimization and differential evolution) graph embedding algorithmic framework, which introduces how to apply gradient-free algorithms on graph embedding problems. Furthermore, the experiments on three large scale real-world network datasets for nodes classification and nodes multi-label classification tasks show that the gradient-free graph embedding algorithms can obtain promising results.