Forecasting stock price movements is challenging due to market volatility and the weak connection between social media sentiment and market indicators. Traditional time-series models struggle with rapidly changing public opinion, while sentiment-based approaches often miss critical technical dynamics. To overcome these issues, we introduce Tweet2Trend, a multimodal framework that combines financial tweets and historical market data. It features a FinBERT-based sentiment encoder for extracting insights from financial tweets and a Bi-LSTM-based technical encoder for analyzing 20-day OHLCV sequences. A cross-attention fusion module aligns these different data types, allowing for effective interaction between sentiment and technical features. Testing on major U.S. stocks, AAPL, TSLA, MSFT, and GOOGL, shows that Tweet2Trend achieves 96.42% accuracy and a 95.83% F1-score, outperforming thirteen strong baseline models. The model performs well even in volatile markets and varying tweet volumes, demonstrating robustness and generalizability. Its lightweight design and low inference latency make Tweet2Trend ideal for real-time trading and decision-support systems, providing a practical solution for financial forecasting.