Emulators are essential for the next stage of surveys with advancing resolution and accuracy. We study various aspects of emulator development for CMB (and other data vectors) by investigating different architectures, training strategies and pre-processing techniques. We demonstrate that, for complicated data vectors like CMB power spectra, a sophisticated model like Transformer or CNN is necessary for high-precision emulation. Clever choices of activation functions, loss functions and rescaling of data vectors also help to make training more efficient. We implement an emulator pipeline interface to Cocoa, a joint architecture of Cobaya and CosmoLike, providing the latest version of our emulators and Cobaya analysis pipelines. In the light of the latest DESI DR2 results, we think about using the Ultra Light Axion (ULA) model to jointly fit CMB and BAO. The axion in the mass spectrum where lg(m_a) ~ -34 to -31.5 behaves like Dark Energy in the early universe, but like Dark Matter in the late universe. Combining such axions and Dark Energy can generate an effective phantom crossing in the effective equation of state. We similarly train emulators for AxiECAMB, an update of AxionCAMB, and run MCMC and profile likelihoods to analyze how well this axion model fits the CMB and BAO datasets in such mass range. Our results show that we can achieve a ΔΧ^2 ~ -7 in CMB+SN+BAO analysis; if we free up \tau and n_s by excluding low ell EE constraints, we can get a ΔΧ^2 ~ -16 (no low ell EE included); if we allow for curvature, we have ΔΧ^2 ~ -12. Only with the case where we release constraints from low ell EE are we able to achieve a comparable improvement as the w0-wa model.