We present a detailed discussion of our new 3D numerical models for the accretion of stellar winds on to \sgra. In our most sophisticated models, we put stellar wind sources on realistic orbits around \sgra, we include recently discovered `slow' winds (v_w 300 km s-1), and we account for optically thin radiative cooling. We test our approach by first modelling only one phase `fast' stellar winds (v_w 1000 km s-1). For stellar wind sources fixed in space, the accretion rate is of the order of \dot M 10-5 \msun yr-1, fluctuates by lt 10%, and is in a good agreement with previous models. In contrast, \dot M decreases by an order of magnitude for wind sources following circular orbits, and fluctuates by 50%. Then we allow a fraction of stars to produce slow winds. Much of these winds cool radiatively after being shocked, forming cold clumps and filaments immersed into the X-ray emitting gas. We investigate two orbital configurations for the stars in this scenario, an isotropic distribution and two rotating discs with perpendicular orientation. The morphology of cold gas is quite sensitive to the orbital distribution of the stars. In both cases, however, most of the accreted gas is hot, producing a quasi steady `floor' in the accretion rate, of the order of 3* 10-6 \msun yr-1, consistent with the values deduced from Chandra observations. The cold gas accretes in intermittent, short but powerful accretion episodes which may give rise to large amplitude variability in the luminosity of \sgra on time scales of tens to hundreds of years. The circularisation radii for the flows are about 103 and 104 Schwarzschild radii, for the one and two-phase wind simulations, respectively, never forming the quasi-spherical accretion flows suggested in some previous work. Our work suggests that, averaged over time scales of hundreds to thousands of years, the radiative and mechanical luminosity of \sgra may be orders of magnitude higher than it is in its current state. Further improvements of the wind accretion modelling of \sgra will rely on improved observational constraints for the wind velocities, mass loss rates and stellar orbits.