# Toto 2.0: Time series forecasting enters the scaling era Source URL: Publisher: Datadog Published: 2026-05-14 Authors: Emaad Khwaja, Gerald Woo, Chris Lettieri, Ameet Talwalkar, David Asker Retrieved: 2026-05-15 ## Article Snapshot This source is a Datadog engineering article announcing Toto 2.0, an open-weights family of time-series forecasting foundation models released on Hugging Face. The article frames Toto 2.0 as a scaling study for forecasting foundation models: model sizes span 4M, 22M, 313M, 1B, and 2.5B parameters, and the reported benchmark curves improve monotonically with scale through 2.5B parameters. The article contrasts Toto 2.0 with Toto 1.0. Toto 1.0 established Datadog's observability-oriented forecasting direction with a 151M-parameter open model. Toto 2.0 keeps the observability framing but expands the release into a model family and emphasizes parameter scaling, inference speed, and benchmark coverage across BOOM, GIFT-Eval, and TIME. ## Official Artifacts - Article: - Source repository: - Hugging Face collection: - Datadog unit-scaling wrapper: ## Model Family The article describes the following Toto 2.0 family members: - Toto-2.0-4m - Toto-2.0-22m - Toto-2.0-313m - Toto-2.0-1B - Toto-2.0-2.5B - Toto-2.0-2.5B-FT, a version fine-tuned on forecasting data - Toto-2.0-FnF, a "Family and Friends" ensemble over Toto 2.0 models and other foundation models The quick-start example in the article installs the package from the `toto2` branch of the official repository and loads `Datadog/Toto-2.0-22m` with `Toto2Model.from_pretrained`. ## Main Claims - Time-series forecasting foundation models can improve predictably with model scale; the article reports no saturation at the 2.5B-parameter endpoint. - Toto 2.0 sets new Datadog-reported best results on BOOM, GIFT-Eval, and TIME. - Toto 2.0 is more parameter-efficient than Toto 1.0; the article states that a much smaller Toto 2.0 variant can match or exceed Toto 1.0 quality. - The model family was trained on observability and synthetic time-series data, with no public forecasting data used in pretraining. - Contiguous patch masking lets Toto 2.0 forecast an entire horizon in one parallel pass, while block decoding remains available for long-horizon stability. ## Training And Inference Notes Toto 2.0 is trained as a forecasting model for time-series dynamics, with a particular focus on observability metrics. The article says the pretraining data mix contains observability metrics and synthetic time series, while deliberately excluding public forecasting datasets to reduce overlap with evaluation benchmarks. The model uses contiguous patch masking during pretraining. Instead of autoregressively revealing one patch at a time, the pretraining task masks a contiguous block and asks the model to predict all masked patches in parallel. At inference time this supports single-pass forecasting for the full prediction horizon. For very long horizons, the article describes block decoding: the model predicts segments sequentially and conditions each segment on the previous segment's median forecast, while using a key-value cache to reduce repeated computation. The article frames this as a speed-versus-stability choice: single-pass prediction is fastest, while block decoding can help keep long-horizon trajectories coherent. ## Reported Benchmark Results On BOOM, the article reports that every Toto 2.0 size outranks every other foundation model by CRPS rank, CRPS, and MASE. Reported CRPS ranks include 3.88 for Toto-2.0-2.5B, 3.96 for Toto-2.0-1B, 4.25 for Toto-2.0-313m, 5.52 for Toto-2.0-22m, 7.17 for Toto-2.0-4m, 6.94 for Toto 1.0, and 7.39 for Chronos-2. On GIFT-Eval's foundation-model leaderboard, the article reports CRPS ranks of 19.5 for Toto-2.0-2.5B, 20.3 for Toto-2.0-1B, 20.5 for Toto-2.0-313m, 25.7 for Toto-2.0-22m, 33.9 for Toto 1.0, 22.1 for PatchTST-FM r1, and 22.4 for Chronos-2. On the full GIFT-Eval leaderboard, the Toto 2.0 Family and Friends ensemble is reported in first place, with a fine-tuned Toto-2.0-2.5B in second place. On TIME, the article reports that Toto-2.0-2.5B, Toto-2.0-313m, and Toto-2.0-1B take the top three positions, while Toto-2.0-22m ranks fifth. The article says every Toto 2.0 size at 22M parameters and above outperforms Toto 1.0 on this benchmark. ## Latency And Long-Horizon Claims The article compares Toto 2.0's single-pass forecasting with the autoregressive style of Toto 1.0 and Chronos-2. It states that a 1024-step forecast can require up to 16 autoregressive steps for Toto 1.0, while Toto 2.0 can produce the same horizon in one forward pass in single-pass mode. It also reports that Toto-2.0-313m has roughly the same latency as Chronos-2 despite being larger, and that for horizons of 2048 or longer even Toto-2.0-2.5B can be faster than Chronos-2 in single-pass mode. For long-horizon forecasting, the article shows synthetic multi-scale examples at 2048, 4096, and 8192 prediction steps. The article's interpretation is that larger Toto 2.0 models preserve temporal structure better than smaller models and prior baselines, but that scale alone does not fully solve very long-horizon forecasting. ## Future Direction The article argues that time-series forecasting still has a gap to close against well-tuned classical methods. It identifies data curation, model scale, benchmark breadth, and better handling of metric-specific structure as future work. The article also sketches a broader observability world-model direction: learning from metrics, traces, logs, service topology, code changes, events, alerts, and text so systems can support proactive incident detection, root cause analysis, counterfactual analysis, simulation, and agent training. In the terminology of this repository, Toto 2.0 is still presented as a passive forecasting model rather than an action-conditioned world model, but the article explicitly points toward future models that could include actions, interventions, and operational context.