AI BriefingDatadog Unveils Time-Series Forecasting Model Toto-2.0
·2026.07.09 20:44
Key point
Datadog announced Toto-2.0, a new time-series forecasting model capable of parallel processing and highly scalable.
Details
Datadog has unveiled Toto-2.0, a new multivariate time-series forecasting model that overcomes the limitations of the existing Toto-1.0. This model has the following key technical features.
- Contiguous Patch Masking (CPM): Instead of the existing autoregressive decoding approach, it uses a single parallel forward pass to improve efficiency.
- Arcsinh Normalization: It compresses extreme spikes in the data while keeping subtle fluctuations clear, optimizing it for handling sparse data.
- NorMuon Optimizer: It handles the sign-valued gradients of the Pinball loss much more effectively than AdamW.
- u-µP Hyperparameter Transfer: Hyperparameters tuned on a proxy model with 10 million parameters can be directly reused across 5 different target model sizes, giving it excellent scalability.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.