Skip to content
Model infoModelsSnapshot: 15 Oct 2026

Nixtla / TimeGPT-2

TimeGPT-2

Standard TimeGPT-2 variant positioned between Mini and Pro for compute cost and forecast accuracy.

Prediction & forecasting
restrictedClosed weightsForecastingAnomaly detection
Official documentation (opens in a new tab)Reviewed 15 Oct 2026
9 of 23 specification fields collected
Context windowNot collectedAwaiting source review
Maximum outputNot collectedAwaiting source review
Input billingSee billing notesNixtla API
Output billingSee billing notesAccess and billing notes below

Specifications

Provider-documented facts. “Not collected” means we have not verified that field yet, not that the feature is unsupported.

Context / input limit
Not collected

Not established for this exact model. Forecasting observations, SDK metadata and time-series usage counters are not text-token limits.

Maximum input tokens
Not collected

Not established for this exact model. Forecasting observations, SDK metadata and time-series usage counters are not text-token limits.

Maximum output
Not collected

Not established for this exact model. Forecasting observations, SDK metadata and time-series usage counters are not text-token limits.

Reasoning
Not collected

Not established for this exact model. Forecasting observations, SDK metadata and time-series usage counters are not text-token limits.

Tool calling
Not collected

Not established for this exact model. Forecasting observations, SDK metadata and time-series usage counters are not text-token limits.

Structured output
Not collected

Not established for this exact model. Forecasting observations, SDK metadata and time-series usage counters are not text-token limits.

Streaming
Not collected

Not established for this exact model. Forecasting observations, SDK metadata and time-series usage counters are not text-token limits.

Batch processing
Not collected

Not established for this exact model. Forecasting observations, SDK metadata and time-series usage counters are not text-token limits.

Fine-tuning
SupportedSource for Fine-tuning (opens in a new tab)

Hosted tuning controls in the shared forecast contract, subject to account entitlement. No private weight download is implied.

Context caching
Not collected

Not established for this exact model. Forecasting observations, SDK metadata and time-series usage counters are not text-token limits.

Parameters
Not collected

Not established for this exact model. Forecasting observations, SDK metadata and time-series usage counters are not text-token limits.

Architecture
Not collected

Not established for this exact model. Forecasting observations, SDK metadata and time-series usage counters are not text-token limits.

Knowledge cutoff
Not collected

Not established for this exact model. Forecasting observations, SDK metadata and time-series usage counters are not text-token limits.

Release date
Not collected

Not established for this exact model. Forecasting observations, SDK metadata and time-series usage counters are not text-token limits.

Languages
Not collected

Not established for this exact model. Forecasting observations, SDK metadata and time-series usage counters are not text-token limits.

Input modalities
Time seriesSource for Input modalities (opens in a new tab)
Output modalities
Numeric forecasts, Prediction intervals, Anomaly flagsSource for Output modalities (opens in a new tab)
Historical input window
Model- and frequency-dependentSource for Historical input window (opens in a new tab)

The SDK queries /model_params for input_size and horizon by model and frequency. No universal numeric input window is asserted.

Forecast horizon
Requested future time steps (h)Source for Forecast horizon (opens in a new tab)

h must be a positive integer. The generic FAQ reports no fixed maximum, but exceeding the native model horizon can reduce accuracy. Account-specific execution limits are not established.

Published price
Custom enterprise quoteSource for Published price (opens in a new tab)

Plans vary API calls, seats, support and deployment. The published 30-day trial is not a permanent free rate.

Usage metering
Forecasting usage; no public numeric rate verifiedSource for Usage metering (opens in a new tab)
Prediction features
Forecasting, optional exogenous features and prediction intervals. The multivariate flag has no effect outside TimeGPT-2.1.Source for Prediction features (opens in a new tab)
Fine-tuning methods
finetune_steps and loss controls in the shared contract. finetune_depth has no effect for TimeGPT-2 family models. Exact account-specific tuning support must be confirmed.Source for Fine-tuning methods (opens in a new tab)

Access & pricing

Prices belong to an endpoint and billing tier. A provider listing does not guarantee access for every account or region.

Nixtla API

Provider source (opens in a new tab)
API model ID
timegpt-2
Access
API key required. TimeGPT-2 variants require account access confirmation. Check /v2/models for entitlement; no live inference performed.
Regions
Account-specific deployment region not established
API endpoint URL
POST https://api.nixtla.io/v2/forecastSource for API endpoint URL (opens in a new tab)
Authentication
Authorization: Bearer API_KEYSource for Authentication (opens in a new tab)

The endpoint schema is authoritative here; older FAQ examples use a different route/header.

Input file formats
JSON numeric targets y and per-series sizes; frequency freq; optional historical X, future X_future and categorical feature indices. SDK accepts pandas/polars long-format dataframes.Source for Input file formats (opens in a new tab)

The dedicated requirements page and current client support polars, despite an older contradictory FAQ answer.

Output formats
Numeric mean array, optional prediction intervals and feature contributions; input_tokens, output_tokens and finetune_tokens usage counters.Source for Output formats (opens in a new tab)
Generation controls
h > 0; prediction interval levels 0 <= level < 100. Future exogenous values must match series count × forecast horizon. No fixed numeric maximum horizon published in the schema.Source for Generation controls (opens in a new tab)
Asynchronous processing
POST /v2/forecast/async; poll /v2/forecast/jobs/{job_id}. Also documented: async tuning, cross-validation, anomaly detection, simulation and explanation.Source for Asynchronous processing (opens in a new tab)
API parameters
Additional native operations: /v2/anomaly_detection, /v2/online_anomaly_detection, /v2/cross_validation, /v2/finetune and /model_params. /v2/explain computes model-agnostic weights, not a separate foundation model.Source for API parameters (opens in a new tab)
Billing scope
Custom enterprise plan. Forecast responses report time-series usage counters, but exact unit conversion and rates require a quote.Source for Billing scope (opens in a new tab)
Request quotas
Not collected

Numeric request, concurrency and payload quotas not established. Plans customize call limits.

Endpoint pricing has not been collected. Check the provider for applicable rates.

Data & enterprise terms

The general privacy notice explicitly excludes business API content and refers to customer agreements. Exact API retention, training use and residency remain contract-specific and unverified.

Read the service policy (opens in a new tab)

Caching

Caching can reuse earlier input processing. Availability, lifetime and billing depend on the endpoint. A cache hit still generates a new response.

01 / INITIAL INPUTProcess & retain

The system processes eligible input and retains reusable state.

02 / LATER REQUESTCheck available state

Reuse depends on matching input, session state and the cache policy.

03 / CACHE HITReuse & generate

Reusing state can reduce repeated work. Any charges follow the billing rules below.

Nixtla API

timegpt-2
Not verified

No reviewed prompt-cache contract or cache tariff. SDK metadata caching and reusing fine-tuned model IDs are different mechanisms.

How it works
Not collected

No reviewed prompt-cache contract or cache tariff. SDK metadata caching and reusing fine-tuned model IDs are different mechanisms.

Minimum reusable input
Not collected

No reviewed prompt-cache contract or cache tariff. SDK metadata caching and reusing fine-tuned model IDs are different mechanisms.

Lifetime & refresh
Not collected

No reviewed prompt-cache contract or cache tariff. SDK metadata caching and reusing fine-tuned model IDs are different mechanisms.

How to enable it
Not collected

No reviewed prompt-cache contract or cache tariff. SDK metadata caching and reusing fine-tuned model IDs are different mechanisms.

Usage fields
Not collected

No reviewed prompt-cache contract or cache tariff. SDK metadata caching and reusing fine-tuned model IDs are different mechanisms.

Billing rules
Not collected

No reviewed prompt-cache contract or cache tariff. SDK metadata caching and reusing fine-tuned model IDs are different mechanisms.

Limits & data boundaries
Not collected

No reviewed prompt-cache contract or cache tariff. SDK metadata caching and reusing fine-tuned model IDs are different mechanisms.

No verified cache rate for this endpoint. An empty rate is not a zero charge.

Caching reviewed 15 Oct 2026

What the measurements say

No benchmark observations have been added for this model. We do not turn a vendor description into a performance score.

Latency and throughput depend on the host, request size, reasoning settings and load. There is no measured speed or global usage figure in this record.

The record behind the model

Why it is included

Nixtla explicitly states that TimeGPT is closed source. Its Apache-2.0 SDK is a client, not a release of the model weights. Customer self-hosting is licensed deployment, not public weight availability.

Classification evidence (opens in a new tab)

Lifecycle dates

Private preview announced14 Oct 2025Source (opens in a new tab)

Sources reviewed