TimesFM Releases 2.5 Time-Series Model Update
Google Research today released TimesFM 2.5, an updated pretrained time-series foundation model that reduces parameter count to 200 million and expands context length to 16,000. The March 31, 2026 update also adds an optional 30M continuous quantile head for up to 1,000-step forecasting, updated inference APIs, and restored covariate (XReg) support. Checkpoints are available on Hugging Face and BigQuery integration is provided.
Key Points
- 1Announces TimesFM 2.5 with 200M parameters and 16k context length for forecasting
- 2Enables long-range temporal pattern capture and efficient inference by cutting model size from 500M to 200M
- 3Provides optional 30M continuous quantile head and API upgrades for robust quantile forecasts and deployment
Scoring Rationale
Significant product update from Google Research with high novelty (16k context, quantile head) and practical usability; credibility is high due to official checkpoints and ICML paper. Score slightly boosted for source authority and freshness, and modestly reduced for ongoing docs/infrastructure updates.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
