Resolved HI kinematics — fitting a tilted-ring model to a galaxy's data cube — underpins much of what we know about rotation curves, disk geometry, and the dark-matter halos that shape them. But because generating a synthetic cube is expensive, established tools largely return a single best-fit solution, fit each ring in isolation, and assume a rigidly axisymmetric disk — hiding the degeneracies that decide whether a result is real. GPU acceleration now removes that bottleneck. I will present FABLE (Flexible Accelerated Bayesian Line-profile Engine), which keeps the entire forward model on the GPU and makes full Bayesian inference of the whole cube routine: it returns joint posteriors rather than point estimates, and fits all rings together. Furthermore, beyond matching established codes to sub-percent accuracy, it can flag non-axisymmetric structure such as bars and lopsidedness and attribute how it biases the inferred dynamics.
I will show validation across beam size, inclination, and rotation-curve shape, and an application to a nearby galaxy in which the full posterior exposed a degeneracy a single best-fit would have hidden — while being candid about where the method is still maturing. Finally, I will walk through the complete procedure end-to-end — from data cube to source extraction, model fitting, and posterior — demonstrated both as an interactive notebook and as a reproducible batch script, so the workflow is something one can pick up and run, and in time apply at scale to the resolved HI samples now arriving from ASKAP, MeerKAT, FAST, and the SKA.
