Toto

by Datadog · 2 versions, 5 parameter counts · Apache 2.0 · HuggingFace · Blog post

forecasting multivariate fine-tunable

Time-series foundation model family from Datadog. Trained on 2 trillion data points of real-world infrastructure, business, and IoT metrics. Toto 2.0 ships in five sizes from 4M to 2.5B parameters, so you can trade off forecast quality against memory and latency.

Quick start

With no flags, routeframe pull grabs the latest version and smallest parameter count:

$routeframe pull toto

That pulls Toto 2.0 / 4M (~16 MB). To see everything available, run:

$routeframe pull toto --list

Versions & sizes

Toto 2.0 default

Current generation. u-µP-scaled decoder with alternating time / variate attention and a quantile output head.

Params	Size (f32)	Pull command
4M default	16 MB	routeframe pull toto
22M	84 MB	routeframe pull toto --params 22m
313M	1.2 GB	routeframe pull toto --params 313m
1B	3.9 GB	routeframe pull toto --params 1b
2.5B	9.1 GB	routeframe pull toto --params 2.5b

Toto 1.0

Original Toto-Open-Base model. Mixture-of-Student-T output head. Supports multivariate inputs and exogenous covariates (via fine-tuning).

Params	Size	Pull command
Open-Base (200M)	289 MB (f16), 577 MB (f32)	routeframe pull toto --version 1.0

Pre-v0.7.0 routeframe clients still default to this version, so existing installations are unaffected by the 2.0 release.

Architecture

Toto 2.0

Architecture: Decoder-only transformer with u-µP scaling
Attention: Alternating time / variate, partial RoPE + xPos on Q/K
Output head: Quantile knots at [0.1, 0.2, ..., 0.9]
Scaler: Patched causal std-mean + asinh
Default context: 512 timesteps
Patch size: 32

Toto 1.0

Architecture: Decoder-only transformer (200M Open-Base)
Output head: Mixture of Student-T (24 components)
Scaler: Causal patch std-mean
Default context: 4096 timesteps
Patch size: 64
Multivariate: Yes (supports exogenous covariates via fine-tuning)

Feature support

Feature	Toto 1.0	Toto 2.0
Zero-shot forecasting	✓	✓
Fine-tuning	✓	×
Exogenous covariates^*	✓	×
Multivariate inputs	✓	✓

^* If you need exogenous covariates, use Toto 1.0.

Run a forecast

$routeframe run toto --input "45,48,52,49,55,58,62,59,64,67" --horizon 4

See the CLI docs for monitoring, exogenous covariates, and fine-tuning.