Aiming to address the problems of difficulty in monitoring tractor-mounted agricultural implements in complex field environments and the excessive amount of model parameters, a lightweight RepViT-based agricultural implements recognition model, tractor-mounted agricultural implements net (TMAInet ), was proposed. Firstly, the self-developed agricultural machinery service platform ‘Agricultural Mechanisation Precision Operation Platform’ was used to collect the datasets of agricultural implements in six working states, and the training set was expanded to 6 627 frames by data enhancement methods such as copy-paste. Secondly, based on the RepVIT network model framework, a convolutional feed-forward module ( CFF) was designed to improve the ability of fine-grained feature extraction at different scales, and an attention mechanism, ECA, was introduced to optimize the model parameter structure and simplify the feature extraction module. Finally, the model was trained by pre-training + fine-tuning (PF) migration learning method and deployed on Jetson nano edge devices. The experimental results showed that the recognition accuracy, F1 score and recall of the TMAInet model reached 99.13% , 98.53 and 98.78% , respectively, by the PF migration learning method. Compared with the original RepVIT model, the recognition accuracy, F1 score and recall were improved by 1.86 percentage points, 3.04 percentage points and 1.95 percentage points, respectively, and the number of parameters was reduced to 7.3 × 10 6 while maintaining 73 f / s at the edge device side. TMAInet was able to accurately and efficiently monitor the common categories of agricultural implements in practical applications, and it can provide a technical reference for the development of unmanned smart farms.