The Transformer architecture and its core attention mechanism form the foundation of Code Language Models (code LMs) and have driven their remarkable progress across a wide range of code intelligence tasks. However, the quadratic complexity inherent in the attention mechanism poses scalability challenges. Recently, subquadratic architectures such as Mamba and Mamba-2 have emerged as compelling alternatives to the Transformer. While they have shown promising results and attracted increasing academic interest, their effectiveness in code intelligence tasks remains unexplored.

To fill this gap, we present the first experimental evaluation of Mamba-based models on two representative code intelligence tasks: line-level code completion and code generation, delving into their effectiveness and efficiency. We begin by evaluating Mamba and Mamba-2 under both full Fine-Tuning (FT) and Parameter-Efficient Fine-Tuning (PEFT) settings. We further pre-train the models on code corpora to boost their capacity in code comprehension. Our results consistently show that Mamba-based models outperform their Transformer-based counterparts. Furthermore, to explore whether the observed superiority stems from the Mamba architecture and to mitigate the influence of varying pre-training datasets, we pre-train CodeGPT, Mamba, and Mamba-2 from scratch on identical code corpora. Our findings reveal that the Mamba-2 block achieves the highest capability in code modeling. Besides, Mamba-based blocks exhibit advantages in terms of memory efficiency. We also evaluate performance in low-resource scenarios and at larger model scales, where Mamba-based models demonstrate consistent robustness. This work provides a comprehensive investigation into Mamba-based models in the context of code intelligence, uncovering their strengths and promising potential for future applications.