White-Basilisk: a hybrid model for code vulnerability detection

White-Basilisk is a hybrid model for code vulnerability detection, and the paper claims it achieves SOTA results with only 200M parameters. The model combines Mamba layers, linear self-attention and MoE, can process ultra-long sequences in a single pass, breaks through the context limits of current LLMs, and comprehensively analyzes large codebases. The model performs robustly on imbalanced real-world datasets, and its computational efficiency supports deployment by organizations of different scales.