How Much Will You Sell Next Week? TimesFM-3 and ChatGPT
🇮🇹 Italiano • 🇪🇸 Español • 🇩🇪 Deutsch • 🇫🇷 Français • 🇧🇷 Português
Have you ever pasted your sales history into ChatGPT to find out how much you'll sell next month? Google's TimesFM-3 is a free download from Hugging Face, but what does its license let you do inside a company? Which alternatives can you ship inside a product you sell? And how do you know whether a forecast holds up on your data? I ran a real backtest to find out.
You’ve probably done it already: paste the last two years of sales into an AI chat and ask how much you’ll sell next quarter. An answer comes back, well written, with numbers in it. But a language model is trained to predict the next word of a text. With numbers, in the best case it writes a small statistical calculation on the fly. In the worst case it hands you a plausible figure. Either way, nobody has measured how wrong that forecast is on your data.
There are models built specifically for numbers. At bit Time Professionals we have been working for years on machine learning models for demand and sales forecasting, system load and inventory trends. This article is the written version of episode 15 of while true do; (in Italian), plus the code and the output of a backtest I ran to write it.
In this article
What’s a time series foundation model?
Until recently, forecasting meant training a custom model for every series: the sales of a store, the load on a server, the energy use of a plant. Time Series Foundation Models are already trained on huge amounts of time series. You run them on your numbers as they are, with no retraining. That’s what’s called zero-shot forecasting.
The idea comes from language models. TimesFM takes the same architecture and adapts it to numbers. Instead of splitting a sentence into words, it splits the series into small blocks of consecutive values (the patches) and learns to predict the next block.
What TimesFM-3 does
Google Research released TimesFM-3 on August 31, 2026. It has 330 million parameters and, according to Google’s blog, it was trained on more than a trillion time points, a mix of real and synthetic data. Google says it ranks first among foundation models on three benchmarks: GIFT-Eval, fev-bench and TIME.
The new feature that matters if you build applications is multivariate forecasting with covariates. Multivariate means the model forecasts several series together and takes into account how they affect each other. Covariates are the external factors behind the numbers: the weather, the holidays, the promotions. TimesFM-3 accepts both the ones you only know for the past and the ones already known for the future, like the dates of the seasonal sales.
Take a home delivery service. The order history alone shows an almost flat curve. Give the model the weather too, and it finds that orders went up on rainy weekends. If the forecast says rain next Saturday, the prediction goes up. So you restock and plan the shifts before you run out of goods.
It also runs on Macs with Apple Silicon, through an MLX backend that doesn’t need PyTorch, and the repository already ships an instructions file for AI agents.
What the license says
The TimesFM code on GitHub is Apache 2.0. The TimesFM-3 weights are a free download from Hugging Face, but under a new license, the TimesFM Non-Commercial License v1.0. Up to version 2.5 the weights were Apache 2.0 too.
This is how the license defines non-commercial use:
“Non-Commercial Purpose” means use for testing, evaluation, or research not tied to commercial gain, production deployment, or revenue generation. This includes internal benchmarking, academic research, and experimentation on private or public datasets, provided the results are not used in commercial decision-making, client deliverables, or paid products/services.
So you can try it and evaluate it, inside a company too, but only as long as the results don’t feed commercial decisions. If you need the forecast to decide how much stock to order, that use isn’t covered, even if the model runs only on your own server and no customer ever sees it. The README puts it plainly: “Commercial or production use of downloaded / self-hosted weights is not permitted”.
For commercial use Google points you to BigQuery. The Google Cloud documentation says that there the use is governed by the Google Cloud terms, which allow it, and that the non-commercial license applies only to the downloaded weights. The BigQuery guides on multivariate forecasting are still marked as Preview, though.
If you want to run the model in-house, or ship it with your software, TimesFM 2.5 is still Apache 2.0. Before you decide, read the whole license and have whoever handles contracts in your company read it too. The link is in the sources at the bottom.
The alternatives
If you want a forecasting engine inside your business application, on ordinary servers with no GPUs and with a license that lets you sell it, there are alternatives. This is the situation as of November 17, 2026, taken from the model cards and the official repositories.
| Model | Made by | Weights license | CPU only | Notes |
|---|---|---|---|---|
| TimesFM-3 | Google Research | non-commercial | not verified | commercial use through BigQuery |
| TimesFM 2.5 | Google Research | Apache 2.0 | not verified | the previous version |
| Timestral | Mistral AI | Modified MIT | yes, stated | 5 to 209 million parameters; free below $20 million in monthly group revenue |
| Chronos-2 | Amazon | Apache 2.0 | yes, tested below | 120 million parameters; related series and covariates together |
| Toto 2.0 | Datadog | Apache 2.0 | not verified | 4 million to 2.5 billion parameters; trained on observability metrics |
| TiRex-2 | NXAI | Apache 2.0 | yes, stated | Docker image for CPU and an ONNX version |
| TimeGPT | Nixtla | weights not downloadable | - | paid API and SDK only |
So if you need to ship the model with your software, today you have Chronos-2, Toto 2.0, TiRex-2 and TimesFM 2.5 with no license constraints, and Timestral as long as your company stays below the threshold.
Be careful with the leaderboards. Google says TimesFM-3 is first, Mistral presents Timestral as better than TimesFM-3, Datadog claims first place for Toto 2.0. Each vendor picks the benchmarks and the metrics, and no leaderboard tells you how wrong the model is on your data.
The backtest, run
To find that out, you run a backtest. You hide the last months of the history from the model, ask it to forecast them, and compare the forecast with what actually happened. Then you run the same test with the most naive forecast there is: “every month equals the same month of the previous year”. If the model can’t even beat that, it’s no use to you on your data.
I did it with Chronos-2, picked because it’s Apache 2.0 and runs on CPU. As data I used monthly car sales in Quebec from 1960 to 1968, 108 months, a public teaching dataset (CSV on GitHub). Python 3.12, chronos-forecasting 2.3.2, CPU-only PyTorch 2.15:
pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install chronos-forecasting pandas
import sys
import pandas as pd
from chronos import BaseChronosPipeline
# Monthly car sales in Quebec, 1960-1968 (108 months)
df = pd.read_csv(sys.argv[1])
df = df.rename(columns={"Month": "timestamp", "Sales": "target"})
df["timestamp"] = pd.to_datetime(df["timestamp"])
df["id"] = "quebec_cars"
MONTHS = 12
history, hidden = df.iloc[:-MONTHS], df.iloc[-MONTHS:]
# Chronos-2 sees only the history and must forecast the 12 hidden months
pipeline = BaseChronosPipeline.from_pretrained("amazon/chronos-2", device_map="cpu")
pred = pipeline.predict_df(history, prediction_length=MONTHS, quantile_levels=[0.1, 0.5, 0.9],
id_column="id", timestamp_column="timestamp", target="target")
chronos = pred["predictions"].to_numpy()
# Naive forecast: each month equals the same month of the previous year
naive = history["target"].iloc[-MONTHS:].to_numpy()
actual = hidden["target"].to_numpy()
print(f"{'month':<8}{'actual':>8}{'naive':>9}{'chronos':>9}")
for m, a, n, c in zip(hidden["timestamp"], actual, naive, chronos):
print(f"{m:%Y-%m} {a:>8.0f}{n:>9.0f}{c:>9.0f}")
def mape(forecast):
return (abs(actual - forecast) / actual).mean() * 100
print(f"\nmean error (MAPE) naive: {mape(naive):.1f}% chronos: {mape(chronos):.1f}%")
This is the whole output:
Loading weights: 100%|##########| 170/170 [00:00<00:00, 187.73it/s]
month actual naive chronos
1968-01 13210 12225 13652
1968-02 14251 11608 14483
1968-03 20139 20985 20187
1968-04 21725 19692 22934
1968-05 26099 24081 23580
1968-06 21084 22114 21527
1968-07 18024 14220 16177
1968-08 16722 13434 13864
1968-09 14385 13598 13231
1968-10 21342 17187 17058
1968-11 17180 16119 17327
1968-12 14577 13713 14852
mean error (MAPE) naive: 10.8% chronos: 6.7%
On this series Chronos-2 is off by 6.7% on average, the naive forecast by 10.8%. Two runs in a row give the same numbers. On my PC, CPU only, one forecast takes about a tenth of a second, measured as the average of five calls after the first load.
Two caveats before you draw conclusions. It’s a single series and a single year. And the dataset is public and widely used, so a model trained on data collected from the web may have seen it already. That’s exactly why the test that counts is the one on your own data, which no model has seen. The code stays the same: you change the CSV and the two columns.
Where to start
Take a series you know well, for example the monthly sales of one product line over the last three years, and run the backtest above. If the model beats the naive forecast, add a covariate you already know matters, like promotions, and see whether the error goes down. If it doesn’t beat it, you stop there, before you’ve spent a whole project on it.
I’m convinced that almost every company, starting with small and medium businesses, already has the right data in-house to use these models without changing the way it works. At bit Time Professionals we offer thirty minutes of free consultation to look at your processes together and see where AI, these models included, gives you a concrete advantage. Write to professionals@bittime.it.
The episode is on while true do; (in Italian). The argument for measurable criteria instead of trust in the tool is the one from episode 14 on Uncle Bob (in Italian), and for what to check in AI-written code there’s the article Zero Errors, Green CI, AI-Written Code: Who Actually Thought It Through?. For questions or corrections: whiletruedopodcast@bittime.it.
Sources
- Google Research, TimesFM-3: a zero-shot foundation model for multivariate forecasting.
- TimesFM repository on GitHub: Apache 2.0 code, Apache 2.0 weights up to 2.5, MLX backend.
- TimesFM-3 on Hugging Face and the text of the TimesFM Non-Commercial License v1.0.
- TimesFM 2.5 on Hugging Face, Apache 2.0 license.
- Google Cloud, TimesFM in BigQuery.
- Mistral AI, Timestral on GitHub, model card and license.
- Amazon Science, Introducing Chronos-2 and the technical report.
- Datadog, Toto 2.0 and the collection on Hugging Face.
- NXAI, TiRex-2 on Hugging Face.
- Nixtla, TimeGPT.
- Dataset: monthly car sales in Quebec, 1960-1968.
Key facts
Article by Daniele Teti (bit Time Professionals), the written version of episode 15 of the podcast "while true do;", about Google's TimesFM-3 and Time Series Foundation Models for demand and sales forecasting in a company.
- TimesFM-3 (Google Research, August 31, 2026): 330 million parameters, trained on more than a trillion time points, stated to rank first among foundation models on GIFT-Eval, fev-bench and TIME; supports past and future covariates and multivariate forecasting.
- License: code Apache 2.0; TimesFM-3 weights under the TimesFM Non-Commercial License v1.0, which excludes using the results for commercial decisions, client deliverables and paid products, even inside a company. Commercial use through BigQuery (Google Cloud terms). Weights up to TimesFM 2.5 remain Apache 2.0.
- Alternatives with Apache 2.0 weights: Chronos-2 (Amazon), Toto 2.0 (Datadog), TiRex-2 (NXAI). Timestral (Mistral AI) uses a Modified MIT license with a threshold of $20 million in monthly revenue. TimeGPT (Nixtla) is available only through an API.
- Backtest run by the author with Chronos-2 on CPU (monthly car sales in Quebec 1960-1968, last 12 months hidden): MAPE 6.7% against 10.8% for the naive seasonal forecast; about 0.1 seconds per forecast. A single series, a public dataset possibly seen in training: the test that counts is on your own data.
- Author: Daniele Teti, bit Time Professionals. Thirty minutes of free consultation: professionals@bittime.it.
Comments