Abstract
Modern networked systems increasingly depend on intelligent models for prediction, control, and scheduling across heterogeneous environments, including video streaming, adaptive bitrate control, and cloud cluster management. However, many existing network foundation models rely on large Transformer-style architectures, which often introduce high inference latency, large model storage requirements, and limited deployability in resource-constrained edge and cloud systems. To address this limitation, we present MambaNetX, a lightweight Mamba2-based sequence modeling framework for multi-modal network intelligence. MambaNetX leverages selective state-space modeling and LoRA-based parameter-efficient fine-tuning to capture long-range temporal dependencies while supporting efficient task adaptation. We evaluate MambaNetX on three representative networking tasks: viewpoint prediction, adaptive bitrate control, and cluster job scheduling. The results show that MambaNetX achieves competitive task-level performance, with an average MAE of 12.5 for video prediction, an average QoE of 96.9 for adaptive bitrate control, and an average JCT of 55.1 for cluster job scheduling. In the system-level evaluation, MambaNetX also demonstrates strong inference efficiency, reaching 421.65 to 7,278 input tokens/s as the input length increases from 32 to 512 tokens. Although the lightweight Mamba2 design may introduce modest task-level tradeoffs compared with larger models, it provides clear benefits in throughput, storage cost, and deployment feasibility. Overall, MambaNetX offers a practical balance between network intelligence, computational efficiency, and real-world deployability.