t0-gamma is our new forecasting foundation model. It has a similar architecture to t0-beta, which our technical report describes in detail. It strengthens the t0 family’s ability to forecast from target history, related series and inputs known ahead of time. Weeks after the release of t0-beta, t0-gamma achieves state-of-the-art performance on public benchmarks. At the time of release, it ranks first among time-series foundation models on GIFT-Eval, fev-bench and TIME.
Number one among time-series foundation models
At the time of release, t0-gamma is the top time-series foundation model on all three public benchmarks we track. On fev-bench, it reaches a 49.42% skill score and a 91.6% win rate. On GIFT-Eval, it scores 0.4520 CRPS and 0.6608 MASE. On TIME, it scores 0.5340 CRPS and 0.6370 MASE.
The same checkpoint is used everywhere, with no task-specific retraining.
GIFT-Eval
CRPS · lower is better
68 eligible models · 97 configurations
- 1t0-gammaThe Forecasting CompanyCRPS 0.4520MASE 0.6608
- 2TimesFM-3GoogleCRPS 0.4557MASE 0.6668
- 3Lingjiang 2.0AlibabaCRPS 0.4565MASE 0.6664
- 4EXAONE-ForecastLG AI ResearchCRPS 0.4600MASE 0.6733
- 5QiYao-MECNU & HuaweiCRPS 0.4608MASE 0.6711
- 9t0-betaThe Forecasting CompanyCRPS 0.4738MASE 0.6865
fev-bench
Skill (%) · higher is better
30 eligible models · 100 tasks
- 1t0-gammaThe Forecasting CompanySkill 49.42Win rate 91.6%
- 2TimesFM-3GoogleSkill 48.99Win rate 86.8%
- 3Chronos-2AmazonSkill 47.28Win rate 81.0%
- 4t0-betaThe Forecasting CompanySkill 46.65Win rate 78.7%
- 5TiRex-2NXAISkill 45.50Win rate 77.6%
TIME
CRPS · lower is better
31 eligible models · 98 tasks
- 1t0-gammaThe Forecasting CompanyCRPS 0.5340MASE 0.6370
- 2QiYao-MECNU & HuaweiCRPS 0.5362MASE 0.6383
- 3TimesFM-3GoogleCRPS 0.5363MASE 0.6398
- 4Toto-2.0-2.5BDatadogCRPS 0.5394MASE 0.6419
- 5Toto-2.0-313mDatadogCRPS 0.5423MASE 0.6444
- 8t0-betaThe Forecasting CompanyCRPS 0.5551MASE 0.6647
fev.leaderboard on the benchmark’s results, where t0-gamma is also first of all 30 entries. Excludes agents, ensembles and fine-tuned systems: on GIFT-Eval several agents score higher, as does TimeRSI, an agent built on TimesFM-3, on TIME. GIFT-Eval’s default view sorts by mean per-configuration rank, where t0-gamma is second to TimesFM-3.The margins at the top are narrow. TimesFM-3 is within 1% of t0-gamma on GIFT-Eval and within half a percent on TIME, and QiYao-M is closer still on TIME. Several agent systems, which orchestrate other models around each task, score higher on GIFT-Eval, and one, TimeRSI, leads TIME. We compare time-series foundation models here, because that is what you download and run.
Head to head on individual tasks
An aggregate score can hide where a model wins. On fev-bench, t0-gamma beats Chronos-2 on 74 of 100 tasks. Against TimesFM-3, it wins 54 tasks overall: an even split on the 42 tasks with covariates and 56.9% of the rest.
How often does t0-gamma win?
Share of tasks with lower forecast loss than each model
| Compared with | With covariates42 tasks | Without covariates58 tasks |
|---|---|---|
| TimesFM-3Google | 50.0% of 42 tasks | 56.9% of 58 tasks |
| Chronos-2Amazon | 66.7% of 42 tasks | 79.3% of 58 tasks |
| t0-betaThe Forecasting Company | 81.0% of 42 tasks | 84.5% of 58 tasks |
| TiRex-2NXAI | 73.8% of 42 tasks | 89.7% of 58 tasks |
| Toto-2.0-2.5BDatadog | 88.1% of 42 tasks | 60.3% of 58 tasks |
| TabPFN-TS-3Prior Labs | 81.0% of 42 tasks | 91.4% of 58 tasks |
| TS-ICLEDF-Lab | 90.5% of 42 tasks | 96.6% of 58 tasks |
| Toto-2.0-22mDatadog | 90.5% of 42 tasks | 75.9% of 58 tasks |
fev.pairwise_comparison from the benchmark’s per-task results as of October 13, 2026, with its missing-result and training-overlap adjustments.The weakest columns point at the next step. The Toto-2.0 models remain competitive on tasks without covariates, and TimesFM-3 holds its own on tasks with them.
A clear step up from t0-beta
t0-gamma improves on t0-beta on every benchmark, with the same leaderboards and the same scoring. On fev-bench, it beats t0-beta on 83 of 100 tasks.
t0-gamma vs t0-beta
- GIFT-Eval−4.6%CRPS, 0.4738 → 0.4520Rank 9 → 1
- fev-bench+2.8 ptsskill, 46.65% → 49.42%Rank 4 → 1
- TIME−3.8%CRPS, 0.5551 → 0.5340Rank 8 → 1
Run it
The t0-gamma weights are on Hugging Face. Our open-source library, tfc-t0, runs them locally in a few lines of Python.
pip install "tfc-t0[torch]>=0.6.0"The quickstart walks you through a first forecast with t0-gamma, with and without covariates.
For managed forecasting, the Retrocast API and Retrocast forecast and backtest your own series without running a model yourself.
For enterprise deployments, larger forecasting workloads, or help evaluating t0-gamma on your own data, talk to us. And if you’d like to build the next one, we’re hiring.