<p>Combining numerical patterns with symbolic representations enables time series forecasting models to more effectively capture multi-scale dynamics, underlying structural trends, and domain-specific interpretability—factors often overlooked by purely numerical approaches. However, existing forecasting tools are generally constrained by fixed-length sliding windows, lack scalability for large-scale temporal data, and often underperform in real-time or zero-sample forecasting tasks. To address these challenges, we introduce BiSTAG-TS, a high-performance dual-stream generative framework designed for symbolic–numerical time series forecasting. The framework first converts continuous time series into LLM-compatible discrete labeled sequences via dynamic discretization, leveraging the sequence modeling and generation capabilities of large pre-trained language models. Simultaneously, it extracts a symbolic event stream representing semantic temporal transitions to enhance interpretability and causal inference. By exploiting parallel autoregressive generation and adaptive fusion between the numerical and symbolic streams, BiSTAG-TS supports highly parallelizable computation suitable for HPC infrastructures, enabling efficient processing of multi-dimensional, high-frequency time series in real-time environments. Moreover, the unified symbolic–numerical representation facilitates large-scale temporal reasoning and multi-domain question answering within a single LLM-based architecture. Extensive experiments on diverse real-world datasets demonstrate that BiSTAG-TS achieves state-of-the-art forecasting accuracy, provides interpretable symbolic insights, and exhibits strong scalability and generalization capabilities under both zero- and few-shot conditions—highlighting its potential for deployment in large-scale, high-performance, and real-time forecasting scenarios.</p>