The FreeFlow architecture introduces a novel approach to optical flow estimation by eliminating task-specific inductive biases such as correlation volumes, feature warping, or iterative refinement. Instead of relying on these predefined heuristics, the researchers developed a hierarchical transformer utilizing a single feed-forward encoder-decoder structure. The model achieves high expressivity through a tripartite attention mechanism: window attention for localized feature processing, shifted-window attention to facilitate cross-window information exchange, and global attention performed at a reduced resolution to capture long-range dependencies. This design allows the model to scale effectively across different capacities, maintaining consistent accuracy improvements from small to large variants. Despite the absence of standard flow-specific components, FreeFlow establishes new state-of-the-art benchmarks, achieving an End-Point Error (EPE) of 0.68 on the Sintel Clean set and 1.48 on the Sintel Final set. Additionally, the model reached 3.23 Fl-all on KITTI-2015 and 3.192 1px on the Spring benchmark. Notably, the architecture remains computationally efficient, supporting memory-efficient inference even at 1080p resolution, proving that a pure transformer-based approach can outperform traditional, bias-heavy pipelines.