Arterial spin labeled (ASL) perfusion MRI allows absolute quantification of cerebral blood flow (CBF) and has found broad applications^1,2. The ability to quantify a purely physiological parameter (CBF) makes ASL MRI an appealing biomarker for clinical research and management. However, ASL perfusion MRI sequences are difficult to calibrate because a reliable perfusion phantom has yet to be developed. Here we describe a prototype perfusion phantom based on 3D printed vessels and mock parenchyma that may allow reliable, ex-vivo assessments of ASL sequences. 3D design was implemented on Blender (Fig. 1). Flow channels were modeled after the brain-supplying arteries³. The phantom was printed on a Dimension Elite (Stratsys) with ABS, and the support material employed to create internal features was removed in a heated bath (70 ^oC) of sodium peroxide solution. A sponge mesh insert was used to fill the parenchymal chamber, simulating a microvascular compartment with exchange of label with bulk tissue water. A Masterflex L/S programmable peristaltic pump (Cole-Parmer) was used to generate continuous flow (also allows pulsatile flow) by pulling water directly from the phantom (Fig. 2). Pulled water is released back into the water reservoir, and before re-entering the phantom, it reaches M_z equilibrium, and air bubbles accumulated in the circuit are removed. Phantom dimensions were calibrated to model physiological CBF (≈ 100 mL/100g/min) when 100 mL/min of flow is applied (Table 1). To validate the phantom performance, 10-pair pCASL⁴-EPI data were acquired at varying labeling durations (LD = 1, 2, 3 sec) or pump rates (PR = 150, 300 mL/min) over a range of PLDs from 100 to 2800 ms, with the following imaging parameters: TE/TR = 12/10000 ms, 4 slices (5.5 mm thickness; 25% gap) covering the chamber, resolution = 4 x 4 mm², matrix = 64 x 64, PF = 6/8, labeling plane distance = 80 mm (positioned at the stem). Water (T₁ = 2836 ms) was used as perfusate. The general kinetic model for continuous ASL⁵ was used to predict the perfusion signal curves: f and τ parameters were set equal to the actual PR and LD values, while the constant terms (i.e. α, λ) were manually adjusted to match the model prediction to the perfusion curves acquired at PR = 300 mL/min & LD = 1 sec and PR = 150 mL/min & LD = 1 sec. While keeping α and λ constant, the model was applied to predict all other perfusion curves. This allowed qualitative assessment of the perfusion signals obtained in the phantom at varying PRs and LDs. The difference signals acquired at multiple LDs show a remarkable similarity to the model predictions (Fig. 3). When LD was increased, the measured signals (circles) correspondingly increased as predicted by the model (dotted lines). After the peaks (PLD > 500 ms), the phantom signals closely followed the estimated T₁ decay, suggesting no occurrence of outflow effects. When PR was decreased, the transit time correspondingly increased (Fig. 4). At PR = 300 mL/min (blue), the peak was observed at PLD = 500 ms, whereas at PR = 150 mL/min (red), the peak was observed at PLD = 1000 ms. In addition, the perfusion signal intensity changed with varying PR as predicted by the model (Fig. 4). Overall, the perfusion signals obtained in the phantom demonstrated good qualitative agreement with the model predictions. We demonstrated that the perfusion signal characteristics of the ASL perfusion phantom are in good qualitative agreement with theoretical predictions based on the general kinetic model⁵. Future work will aim to characterize α in the stem and the effects of non-continuous flow patterns on labeling efficiency.