This article tackles the problem of online state estimation and parameter identification for SMA-actuated flexible structures intended to be precisely controlled within a multivariable adaptive model-based framework. Using a joint state-parameter formulation, a non-linear recursive scheme has been developed that is capable of simultaneously producing online estimates of the states and the uncertain parameters from the noisy measurements at hand. The scheme employs an embedded model derived from reduced-order finite element modeling of the structure and phenomenological SMA modeling to incorporate the whole available knowledge about the nature of the system (non-linearity and hysteresis) and its uncertainties (uncertain parameters and stochastic noises). The unscented Kalman filtering algorithm is utilized to improve accuracy and simplify the implementation. The numerical functionality of the proposed scheme is validated via a simulation example; while its practical versatility is challenged by an experimental case study involving the SMA-actuated flexible tail of a bio-inspired ornithopter. Results demonstrate successfulness of the scheme for online estimation and identification purposes, as well as promising applicability to adaptive nonlinear model-based control.