Traditional system identification usually follows two steps: first, training data are used to estimate a model of the target system; then the model predicts outputs for new inputs. Recent in‑context learning approaches bypass explicit model estimation by training a meta‑model that maps training data and query inputs directly to outputs, enabling zero‑shot prediction. The meta‑model is trained offline on large synthetic datasets to emulate a system‑identification algorithm.\ \ This paper advances the meta‑learning framework in three ways. First, it proposes a unified probabilistic meta‑identification architecture that combines a Transformer encoder‑decoder with mechanisms tailored for long training sequences, allowing calibrated predictions over extended horizons. Second, it shows that a pre‑trained meta‑model can serve as an uncertainty‑aware synthetic data generator for classical system identification, supplying task‑specific data that regularize parametric models when real data are scarce. Third, it conducts systematic evaluations on standard nonlinear system‑identification benchmarks and a challenging out‑of‑distribution scenario, demonstrating competitive accuracy with state‑of‑the‑art methods and strong generalization, even when only input‑only adaptation is used to handle significant distribution shifts.\ \ Review: The study effectively transfers the contextual learning capabilities of large models to the control domain, offering a promising solution for data‑limited identification tasks. Its dual benefits of probabilistic forecasting and synthetic data generation open new research directions.