With the rapid development of the cryptocurrency market, accurately forecasting price trends has become a critical research topic. In recent years, the surge of social media has made sentiment analysis a prominent area of study. This paper proposes an innovative hybrid model that combines Convolutional Neural Networks (CNN), Bidirectional Long Short-Term Memory (Bi-LSTM) networks, and a multi-head attention mechanism to predict short-term cryptocurrency price movements, incorporating news sentiment as a key input feature. Specifically, the one-dimensional convolutional layer (Conv1D) is employed to extract local features from time series data, the Bi-LSTM network captures long-term temporal dependencies, and the multi-head attention mechanism learns various attention representations to uncover relationships across multiple dimensions. This architecture allows the model to establish long-range dependencies between different segments of the time series and to identify latent patterns within various time windows. The experimental datasets consist of historical cryptocurrency news collected from platforms such as Cryptonews.com, Cryptopotato.com, and Cointelegraph, along with historical hourly price data including open, high, low, and close (OHLC) values. The model uses data from the preceding 30 h to predict the price movement in the next hour. Seven major cryptocurrencies—Bitcoin (BTC), Ethereum (ETH), BNB (BNB), Ripple (XRP), Dogecoin (DOGE), Solana (SOL), and Cardano (ADA)—are selected for analysis. Among these, the lowest prediction accuracy was achieved for ADA at 73.31%, while XRP had the highest accuracy at 78.65%. Overall, the results demonstrate that the proposed model achieves robust predictive performance across major cryptocurrencies.