TOKI 1.0
TIME SERIES · FM

toki 1.0
Zero-shot forecasting.

Forecast multivariate time series in a single API call — no training required. Built for enterprise inference.

Benchmark numbers on this page are placeholders.

WHY

Why a foundation model for time series?

Classical forecasting requires training a separate model per dataset and tuning hyperparameters by hand. Lead times are long and maintenance costs are recurring.

Cold-start regimes — new sites, new SKUs — are structurally difficult for models that have not seen comparable history.

Time series foundation models are pre-trained on diverse data and applied zero-shot to any task. toki 1.0 brings this approach to production quality.

FEATURES

What sets it apart

01

Zero-shot

Use the pre-trained model as-is. No task-specific fine-tuning — call the API and get predictions back.

02

Multivariate

Forecast multiple series jointly. Cross-series dependencies are captured for sharper results than per-series models.

03

Covariates

Pass numerical and categorical covariates — holidays, weather, campaign types — alongside the target series.

04

Cold-start

Generate reasonable forecasts for new series with little history by leveraging patterns learned at scale.

API

Forecast in a single request

No training required — call the API and get predictions back. Available via SDK or REST.

curl
curl -X POST https://api.toki.neurogica.com/v1/forecast \
  -H "Authorization: Bearer $TOKI_API_KEY" \
  -d '{
    "series": [...],
    "horizon": 24
  }'
Python
from toki import Client

client = Client(api_key="...")
forecast = client.forecast(
    series=df,
    horizon=24,
)
ARCHITECTURE

How the model works

Inputs are normalized and patched, processed by a Transformer that captures intra- and inter-series structure, and emitted as a probabilistic forecast.

  1. 01 / Tokenize
    Tokenize
    Range normalization and numeric tokenization
  2. 02 / Patch
    Patch
    Split series into fixed-length patches
  3. 03 / Transformer
    Transformer
    Learn intra- and inter-series dependencies
  4. 04 / Forecast
    Forecast
    Emit quantile-aware forecast distributions
BENCHMARKS

Outperforms baselines, zero-shot

DEMO

Evaluated against major open-source time series foundation models on axes equivalent to fev-bench and GIFT-Eval.

fev-bench win-rate

fev-bench win-ratetoki 1.00.72Baseline A0.61Baseline B0.58Baseline C0.49

Skill score (higher is better). All numbers are placeholders.

GIFT-Eval ranking

Probabilistic forecast
RankModelScore
#1toki 1.00.81
#2Baseline A0.74
#3Baseline B0.71
#4Baseline C0.66
Point forecast
RankModelScore
#1toki 1.00.78
#2Baseline A0.70
#3Baseline B0.68
#4Baseline C0.61
CONTACT

Considering a PoC?

Tell us about your use case, data scale, and requirements. Our engineering team will reach out from there.