Background and Objective: Quantifying blood flow in embryonic vasculature from optical coherence tomography (OCT) images is hindered by severe speckle noise, low contrast and the absence of ground-truth motion labels. Particle image velocimetry and generic deep learning optical flow models are sensitive to speckle, suffer from domain mismatch and depend on dataset-specific preprocessing. This study aimed to develop a weakly supervised deep learning framework for dense flow estimation in noisy embryonic OCT image sequences.,Methods: We propose OCT-RAFT, which couples a pretrained Recurrent All-Pairs Field Transforms (RAFT) backbone with a lightweight U-Net refinement module. Because annotated flow data are unavailable, the model is trained with Farneback-derived optical flow as a weak supervisory signal and a center-weighted spatial loss mask that emphasizes high-motion vascular regions; the loss combines masked L1 flow, relative angle and spatial smoothness terms. The model was trained on seven OCT sequences from seven chick embryos using a single standardized preprocessing pipeline. It was evaluated on an in vitro microfluidic phantom with a known flow rate and on 15 independently acquired in vivo chick embryos at Hamburger–Hamilton stages 18, 20 and 22 (n = 5 per stage), and compared with particle image velocimetry, Farneback dense optical flow, SPyNet, pretrained RAFT and fine-tuned RAFT. Results were analyzed descriptively using means, standard deviations and 95% confidence intervals.,Results: In vitro, OCT-RAFT velocities agreed with continuity-based theoretical predictions within 8.1% at the channel inlet and 8.3% at the widest section, with inter-run coefficients of variation of 2.1% and 6.0% for the inlet and channel-averaged velocities. In vivo, OCT-RAFT yielded mean velocity magnitudes of 0.11 ± 0.06, 0.19 ± 0.11 and 0.48 ± 0.15 mm/s at stages 18, 20 and 22, compared with 0.32 ± 0.30, 0.81 ± 0.44 and 3.30 ± 1.70 mm/s for particle image velocimetry, and produced spatially coherent flow fields with fewer background artifacts than the baseline methods.,Conclusions: OCT-RAFT provides a weakly supervised framework for quantitative blood flow analysis in embryonic OCT imaging without dataset-specific tuning, and its design may transfer to other speckle-rich biomedical imaging modalities.