Few-shot font generation targets to generate characters of a specific font with only a small number of seen characters available as references. In recent studies, a common approach is representing a target font by a group of basis fonts with appropriate mixing weights, and the style features of a generated target character could be controlled by selecting different basis fonts and adjusting their mixing weights. Fonts with similar styles encoded from the same content are mapped into the same cluster in a latent space to learn a group of clustering centers as basis fonts to enable few-shot font generation via a combination of basis fonts. However, recent studies cluster basis fonts according to the Euclidean distances between font content features in a high dimensional space, which would fail to report the ratio between the nearest and farthest neighbors to a given target, driving to inaccurate selection of basis fonts. Moreover, the computed mixing weights might be over balanced to lose the domination of key basis font(s) during font combination, resulting in unstable generating performance. To address these issues, this paper proposes HG-Font, which maps font features onto a unit hypersphere and uses angle-based cosine similarity as the similarity metric, replacing distance measures. This effectively alleviates the challenges of similarity computation in high-dimensional space. Moreover, we design a dynamic weight gating module that adaptively adjusts the fusion weights according to the input features, collaboratively optimizing the weight distribution and significantly mitigating the averaging problem. Experimental results demonstrate that HG-Font achieves good performance across multiple evaluation metrics, particularly under complex font structures and extremely few-shot settings, showing superior generation capability and stability.