Smart contracts are the backbone of decentralized applications, enabling trustless execution of financial and logical agreements on the blockchain. However, their immutability and complexity make them particularly susceptible to security vulnerabilities, which have led to numerous high-profile exploits. Although traditional static analysis tools and binary classifiers can detect some vulnerabilities, they often fall short in identifying multiple types simultaneously or understanding nuanced patterns in contract logic. In this work, we present a transformer-based approach to multiclass vulnerability detection in Solidity smart contracts. Leveraging CodeBERT, a pre-trained language model for source code, we fine-tuned CodeBERT on a dataset of 5,000+ Solidity contracts from ScrawID, manually annotated with 12 vulnerability types (e.g., reentrancy, integer overflow). Contracts are multi-labeled across 12 categories, reflecting real-world scenarios where multiple flaws coexist, enabling a more granular and comprehensive analysis. Our model achieves 98% accuracy evaluated on a held-out test set of 1,000 contracts, with per-class F1 scores exceeding 95% for common vulnerabilities like reentrancy, with high precision and recall across multiple vulnerability categories. We demonstrate that transformer-based architectures not only outperform traditional techniques but also scale effectively to real-world contract analysis. This research highlights the potential of deep learning in enhancing smart contract security and moves us closer to automating the auditing process with greater accuracy and efficiency.