GIFT-Eval Compares Timestral, t0, Tabby and Morpheus-2.0

Between 10 September and 16 October 2026, Huawei’s Noah’s Ark Lab in Paris, The Forecasting Company, Mistral and JPMorganChase released or highlighted time-series foundation models Tabby, t0-gamma, Timestral and Morpheus-2.0. These systems aim to forecast new sequences, such as hourly electricity demand, without training a separate model for every series. The releases reflect rapid field growth: arXiv references to time-series foundation models rose from five in 2023 to 273 by 30 October 2026, while Amazon’s Chronos-2 recorded 11 million Hugging Face downloads in 30 days. The four models share a common technical approach. They scale numerical inputs, group historical values into patches, use transformer attention, produce probabilistic quantile forecasts and rely heavily on synthetic training data. Their main differences lie in how they read related series and generate forecasts. Tabby forecasts each series independently and cannot use known drivers such as temperature forecasts. Timestral and t0 use selected cross-series attention layers, allowing targets to read known future covariates while preventing future target values from leaking into the forecast. Morpheus-2.0 uses dual-axis attention, mixing time and variable dimensions in every layer. Timestral appends future forecast tokens, t0 predicts the next patch, and Tabby and Morpheus fill masked future positions. Training data and procedures appear especially important: Morpheus’s ablations show large gains from synthetic data and curriculum training, while t0-gamma improves over t0-beta despite an unchanged architecture and only small gains from inference-time adjustments. On Salesforce’s GIFT-Eval benchmark, covering 97 tasks from 55 public datasets, Timestral-209M recorded the best probabilistic CRPS at 0.446, ahead of t0-gamma at 0.452 and Morpheus-2.0 Multiscale at 0.455. Task-level resampling indicates the top three cannot be statistically separated by average score. Morpheus led MASE at 0.646, showing that rankings can change when users prioritize median forecasts over full uncertainty. Domain results also diverged: Morpheus was strongest in energy and transport, where related sensors or variables are common, while Timestral and t0 were closer overall and generally led in web, nature and healthcare tasks. Reproduction with public weights matched official scores on an A10G GPU, but important caveats remain. Morpheus’s headline result comes from a 413M-parameter Multiscale ensemble without public weights, Tabby’s leaderboard entry is prompt-tuned and requires a long minimum forecast span to match its published score, and t0’s notebook includes inference adjustments. Licensing also matters: t0 offers Apache-2.0 weights, while Timestral has a modified MIT licence with a revenue cap and Tabby is non-commercial. For practitioners, the key conclusion is that overall leaderboard order should not decide model choice. Covariate support, domain characteristics, forecast horizon, calibration, compute cost and commercial terms may matter more. Smaller models can be competitive: Timestral-28M is close to the 209M version and is faster on CPUs. A more revealing next step would be testing energy forecasts with real weather covariates, because history-only benchmarks do not fully show which models exploit external drivers effectively.