
Deploy Chronos | (Just Updated) Time-Series Forecasting API, Zero-Shot Quantile Forecasts, API Key Pre-Set
Amazon Chronos-2 time-series forecasting API, zero-shot quantiles, CPU
chronos
Just deployed
Deploy and Host Chronos Time-Series Forecasting API on Railway
Amazon Chronos-2 is an open pretrained time-series foundation model: it forecasts a series it has never seen, with no training and no per-series fitting, and returns quantiles instead of one number. This template serves Chronos-2 and the two Chronos-Bolt sizes behind one authenticated JSON API: post a history, get back a median and quantile forecast, optionally with past and future covariates. The weights are already inside the image, an API key is generated for you at deploy, and everything runs on CPU.
About Hosting Chronos
This template runs chronos-forecasting 2.3.2 (Amazon Science, the newest release at build time) with PyTorch on CPU behind a small FastAPI server. POST /v1/forecast takes series (each with values, oldest first, null for a gap), prediction_length, and optionally quantiles, model (chronos-2, chronos-bolt-small or chronos-bolt-base), start plus freq for timestamps, and, for Chronos-2, past_covariates, future_covariates and cross_learning. The response carries median, a quantiles map and timestamps per series. GET /v1/models lists the models. Chronos-Bolt forecasts at most 64 steps and answers 422 above that; Chronos-2 forecasts up to 1024.
The weights are baked into the image, so the first request is answered from disk and no model download hides inside a request that Railway's edge would abandon after five minutes. The service sizes its PyTorch thread pool from the container's own CPU quota instead of the host's core count, handles one request at a time so memory stays bounded, and caps one request at 2,000,000 values in total.
Measured on Railway, 8 vCPU, 8 GB container, Chronos-2 unless stated: one series of 512 points and a 24-step horizon took 0.06 s (Chronos-Bolt small 0.09 s, Bolt base 0.17 s); 20 series of 1000 points took 0.53 s; 100 series of 2000 points and a 96-step horizon took 6.5 s; 256 series of 4000 points and a 256-step horizon took 35.3 s; and the largest request the service accepts, 100 series of 20,000 points (2.0 million values), took 26.8 s. Every one finished far inside the five minute edge limit. The model loads in about 3.5 s after the container starts, and the process peak resident memory after those requests read 4.5 GB (peak_rss_mb on /health).
Agent contract
- Public URL:
https://RAILWAY_PUBLIC_DOMAIN(the generated domain of thechronosservice) - Private URL:
http://chronos.railway.internal:8080 - Credentials: the
API_KEYvariable of thechronosservice - Auth: send
Authorization: Bearer $API_KEYon every route exceptGET /health; a call without the key or with a wrong one answers401 - Health:
GET /healthreturns 200 once the model is loaded (503 while it loads)
# needs: RAILWAY_PUBLIC_DOMAIN, API_KEY (a variable of the chronos service); also curl.
# signup: none - the API has no user accounts, only the API_KEY bearer
R="https://$RAILWAY_PUBLIC_DOMAIN"
n=0; until [ "$(curl -s -o /dev/null -w '%{http_code}' "$R/health")" = 200 ]; do n=$((n+1)); [ "$n" -le 60 ] || break; sleep 5; done
curl -s -o /dev/null -w "health: %{http_code}\n" "$R/health"
# 8 weeks of a sales-like weekly pattern: 10, 20 ... 70, repeated
V="10,20,30,40,50,60,70"; for w in 1 2 3 4 5 6 7; do V="$V,10,20,30,40,50,60,70"; done
B="{\"series\":[{\"id\":\"sales\",\"values\":[$V]}],\"prediction_length\":7}"
curl -s -o /dev/null -w "write without key: %{http_code}\n" -X POST "$R/v1/forecast" -H "Content-Type: application/json" -d "$B"
curl -s -o /dev/null -w "write wrong key: %{http_code}\n" -X POST "$R/v1/forecast" -H "Content-Type: application/json" -H "Authorization: Bearer not-the-key" -d "$B"
curl -s -X POST "$R/v1/forecast" -H "Content-Type: application/json" -H "Authorization: Bearer $API_KEY" -d "$B" \
| grep -o '"median":\[[^]]*\]' | head -1
A fresh deploy prints health: 200, write without key: 401, write wrong key: 401 and a "median" list close to 10, 20, 30, 40, 50, 60, 70, the weekly pattern continued (the last run printed 8.77, 19.67, 30.29, 40.81, 50.31, 59.91, 68.77). To forecast with a known-in-advance covariate such as a promotion calendar, and to label the output with dates:
curl -X POST "https://$RAILWAY_PUBLIC_DOMAIN/v1/forecast" \
-H "Authorization: Bearer $API_KEY" -H "Content-Type: application/json" \
-d '{"series":[{"id":"store-1","values":[...],"past_covariates":{"promo":[...]},"future_covariates":{"promo":[...]}}],
"prediction_length":14,"start":"2026-11-20","freq":"D","quantiles":[0.1,0.5,0.9]}'
start is the timestamp of the last history point. past_covariates have the same length as values; future_covariates have prediction_length values and must also appear in past_covariates. A malformed request answers 422 with the reason, and a body over MAX_BODY_MB answers 413.
Why Deploy Chronos Time-Series Forecasting API on Railway?
Railway is a singular platform to deploy your infrastructure stack. Railway will host your infrastructure so you don't have to deal with configuration, while allowing you to vertically and horizontally scale it.
By deploying Chronos on Railway, you are one step closer to supporting a complete full-stack application with minimal burden. Host your servers, databases, AI agents, and more on Railway.
- No training step — a zero-shot model forecasts a new series from its history alone, so there is no per-series fitting job to run or schedule.
- Quantiles, not a single line — every forecast comes with the 10th and 90th percentile (or any levels you ask for), which is what stock and capacity decisions need.
- Your data stays in your container — sales, sensor and usage history are forecast by your own service, not sent to a hosted forecasting API.
- Fast on plain CPU — a typical request takes well under a second on 8 vCPU, with no GPU bill and no per-call fee.
- Private by default — a generated bearer key guards every route but the health check, and the container refuses to start without one.
Common Use Cases
- Demand and capacity planning — forecast sales, orders, traffic or queue depth from the history in your database.
- Metric forecasting for agents — give an LLM agent a forecasting tool it can call with one authenticated request, instead of asking it to extrapolate numbers.
- Anomaly baselines — compare live values to the 10th to 90th percentile band the model predicted for them.
Dependencies for Chronos Hosting
- A service with 6 GB of RAM or more (peak resident memory measured at 4.5 GB after requests up to the service's limits). Free and Trial plans cannot run it.
- No volume: the service keeps no state and the models are in the image.
Deployment Dependencies
- amazon-science/chronos-forecasting — the inference library, Apache-2.0, version 2.3.2
- amazon/chronos-2, amazon/chronos-bolt-small and amazon/chronos-bolt-base — the models
- bon5co/chronos-railway — the Railway wrapper image this template deploys, Apache-2.0
- Weights licence: the Hugging Face model card of
amazon/chronos-2declareslicense: apache-2.0, and the Hugging Face API listslicense:apache-2.0foramazon/chronos-2,amazon/chronos-bolt-smallandamazon/chronos-bolt-base. Apache-2.0 allows commercial use and requires keeping the licence and attribution notices. Attribution: Chronos-2 and Chronos-Bolt, copyright Amazon Science, licensed Apache-2.0.
Template Content

