Google's TimesFM-2.5: 200M-Parameter Forecasting Model with 16K Context Tops GIFT-Eval
Google Research has published TimesFM-2.5, a compact but capable time-series foundation model designed for real-world forecasting workloads. The checkpoint is available on Hugging Face and brings a mix of efficiency and extended context that improves zero-shot forecasting performance.
What TimesFM-2.5 brings
TimesFM-2.5 is a decoder-only model with 200 million parameters and native probabilistic forecasting support. Its most notable technical upgrade is a maximum context length of 16,384 time steps, enabling the model to see much longer histories in a single forward pass. An optional 30M-parameter quantile head provides continuous quantile forecasts up to a 1,000-step horizon.
Key differences from v2.0
- Parameters: reduced to 200M from 500M, improving compute efficiency and inference speed.
- Context length: increased to 16,384 points from 2,048, allowing multi-seasonal and low-frequency patterns to be modeled without stitching.
- Inputs and inference: no explicit frequency indicator is required; new inference flags include flip-invariance, positivity inference, and a quantile-crossing fix.
- Roadmap: a Flax implementation is planned for faster inference, covariates support will return, and documentation is being expanded.
Why longer context matters
A 16K context length allows the model to capture multi-seasonal structure, regime shifts, and long-range dependencies in a single pass. This reduces the need for complex preprocessing, tiling, or hierarchical stitching, and improves stability in domains where history far exceeds the forecast horizon, such as energy load forecasting or retail demand planning.
Benchmark performance and availability
On the GIFT-Eval benchmark, TimesFM-2.5 now leads zero-shot foundation models on both MASE (point accuracy) and CRPS (probabilistic accuracy). The model is live on Hugging Face and positioned as production-ready thanks to its compact size, probabilistic forecasting features, and efficient design.
Research context and practical implications
TimesFM builds on the thesis introduced in the ICML 2024 paper and accompanying research blog: a single, decoder-only foundation model can serve diverse forecasting tasks. GIFT-Eval, developed by Salesforce, helps standardize evaluation across frequencies, horizons, and univariate versus multivariate regimes, making cross-model comparisons clearer for practitioners.
Practical takeaway
TimesFM-2.5 demonstrates a shift from proof-of-concept to practical tooling for time-series forecasting: it halves parameter count relative to 2.0, extends context significantly, and leads benchmarks in zero-shot settings. With Hugging Face availability and planned integrations, it is primed to accelerate adoption of zero-shot forecasting in production pipelines.
Resources
Check the model card on Hugging Face, the repository, benchmark details and the original paper for more technical depth and usage examples. The authors also provide tutorials, code, notebooks, and community links for continued learning and adoption.