State-of-the-art adaptive bit rate (ABR) strategies, either learning-based or condition-wised approaches, resort to pre-training one or more ABR decision models offline in a simulator under a limited number of collected network traces. However, these existing ABR approaches struggle to generalize well to current heterogeneous network environments, especially unseen network conditions. In this paper, we propose OLNC, a network condition-wised hybrid ABR system, which combines an offline multi-model switching scheme and an additional online learning mechanism to improve its generalization capability to different network conditions. In the offline stage, OLNC trains an ABR model for each network condition and dynamically conducts model switching according to the identified network conditions. In the online stage, OLNC introduces an online learning mechanism, which automatically detects new network conditions and performs online learning to fine-tune existing models without redesigning or retraining them to guarantee satisfactory QoE under unseen network conditions. Both trace-driven and real-world environmental results show that OLNC outperforms state-of-the-art schemes, with QoE improvements of 33.1%-49.5% across a wide range of network conditions.