Top Highlights
-
A shared recipe is emerging: all four models use scaling, patches, transformers, quantile forecasts and lots of synthetic training data.
-
The key difference is what they can “see”: Tabby forecasts each series alone; Timestral, t0 and Morpheus can use related series—potentially valuable covariates like tomorrow’s weather.
-
No clear leaderboard winner: Timestral leads on probabilistic accuracy, Morpheus on median forecasts, but the top models’ overall differences aren’t statistically decisive. Training seems to explain more gains than inference tweaks.
-
Choose for your real-world needs: test models on your own data, covariates and forecast horizon—and weigh speed, access and licensing alongside accuracy.
One forecasting goal, three architectural choices
Timestral, t0, Tabby and Morpheus-2.0 all use transformer-based models, patches and predicted ranges to forecast time-series data. Yet they differ in what information they can use.
Tabby forecasts each series on its own. It cannot directly use a separate series, such as tomorrow’s temperature, to predict electricity demand. Timestral and t0 allow selected layers to share information across series. Morpheus does so in every layer, while limiting attention to control memory.
That flexibility may help when outside drivers matter. However, GIFT-Eval mostly tests forecasts based on each series’ own history. Its scores therefore offer limited evidence about how well models use known future inputs.
Close scores, different strengths
On GIFT-Eval, Timestral-209M leads on probabilistic CRPS with a score of 0.446. T0-gamma follows at 0.452, and Morpheus-2.0 scores 0.455. Lower scores indicate better results. Yet statistical comparisons cannot reliably separate these three models.
The ranking changes for point forecasts. Morpheus leads on MASE, with 0.646, compared with Timestral’s 0.658. The result also reflects different model setups: Morpheus’ leaderboard score comes from a five-model ensemble, while Tabby’s leading entry uses tuned prompt vectors.
So, the leaderboard offers a useful starting point, not a universal verdict. A model’s performance may shift with the task, metric and forecast horizon.
Deployment choices shape adoption
For practical use, test models on data that resembles the real job. A utility forecasting demand alongside weather forecasts should compare models with and without those inputs. It should also check speed, memory needs and forecast uncertainty.
Access and licensing matter, too. T0 offers Apache-2.0 weights. Timestral’s modified MIT license includes a commercial revenue restriction, while Tabby’s weights carry a noncommercial license. Morpheus’ public weights were not available when the comparison was prepared.
Meanwhile, smaller models can offer strong results at lower cost. Timestral-28M trails its larger counterpart by just 0.015 CRPS. That makes local testing—and not a small leaderboard gap—the better guide to adoption.
Expand Your Tech Knowledge
Stay informed on the revolutionary breakthroughs in Quantum Computing research.
Discover archived knowledge and digital history on the Internet Archive.
AITechV1
