We introduce a phase‑aware self‑supervised framework (MSPR‑GRE) for accelerated high‑resolution multi‑echo gradient‑echo (mGRE) MRI that preserves anatomical detail and complex echo information for quantitative mapping. MSPR‑GRE integrates progressive k‑space expansion with phase estimation and refinement. Sequentially unrolled networks first reconstruct complex multi‑echo images, then update echo‑dependent phase priors. For acquisitions with at least four echoes, a relaxation‑informed temporal subspace is employed; shorter echo trains are reconstructed directly in the image domain. Training uses only held‑out k‑space samples, eliminating the need for fully sampled targets. Evaluation was performed on 9.4 T ex‑vivo mouse‑brain data (30 µm eight‑echo, 18 µm dual‑echo) and retrospectively undersampled 7 T AHEAD human‑brain data, with supervised ablations for comparison. MSPR‑GRE reduced artifacts in regions of rapid phase variation and retained fine mouse‑brain anatomy. Across five AHEAD test sets at acceleration factors R=4, 8, and 12, it consistently improved magnitude images, T2* maps, and QSM metrics, especially at higher accelerations. At R=12, the fourth‑echo magnitude NRMSE was $12.09\pm0.60\%$, compared to $21.55\pm1.12\%$ for SSDU. Ablation studies highlighted the benefit of phase refinement, and supervised experiments confirmed the advantage of multi‑stage phase‑aware reconstruction. In conclusion, MSPR‑GRE enhances accelerated mGRE reconstruction and quantitative‑map agreement without fully sampled training targets, though further validation of generalizability and phase‑estimation accuracy is required.\n\nReview: The study cleverly combines self‑supervised learning with phase priors to achieve high‑acceleration MRI reconstruction while maintaining quantitative fidelity, offering a promising direction for future clinical and research applications.