Is there an "AlphaGo" for Xiangqi?
No, but there is px0, which uses the same algorithm (website: px0.org). However, in Xiangqi, CPU engines are currently still stronger than GPU engines.
AlphaGo and AlphaZero from years ago used a different search algorithm, built to work with the real centerpiece: a convolutional neural network architecture. These engines mainly run on a graphics card (GPU) or TPU. The search algorithms of strong Xiangqi engines, on the other hand, need only a CPU and no graphics card.
The networks of GPU engines are large enough to evaluate positions more accurately. Unlike NNUE, the network of a GPU engine is not just an evaluation network; it also contains a policy network, which decides which moves to search. This is similar to a human looking at a position and needing to pick a move to start calculating: if this "intuition" is accurate enough, it saves a great deal of time. But this large architecture makes GPU engines more dependent on graphics cards and MCTS search (in fact, GPU engines for board games are no longer Monte Carlo in the narrow sense). At present, GPU engines have still not surpassed CPU engines in Xiangqi or chess.
The Alpha series is not open source. KataGo and Leela for Go and lc0 for chess were all reproduced from its papers, while for Xiangqi there are px0 and ggz, which are clones of lc0.
Today's top Go, chess and Shogi engines are already far stronger than AlphaZero back then.
For GPU engines, whether the upper limit differs between iterative training starting from random play and from human data remains an open question. Starting from random play may just be a gimmick, since a0 and ag0 also differ in other respects.
Note
AlphaCat (阿尔法猫) uses a different algorithm from AlphaGo, namely ab pruning. It merely has a similar name.