What is TimesFM?
TimesFM (Time Series Foundation Model) is a decoder-only Transformer from Google Research that forecasts numeric series zero-shot: you pass in past values and it returns a forecast plus nine quantiles, without fitting a model per series. The original paper appeared at ICML 2024. The repo holds the code, examples, a LoRA fine-tuning recipe and an agent skill file, with PyTorch, Flax and an MLX backend for Apple silicon.
TimesFM 3.0, released in late August 2026, has 330 million parameters and adds multivariate forecasting and covariates known only in the past or also in the future. Google's blog says it was trained on more than 1 trillion time points and ranks first among foundation models on GIFT-Eval, fev-bench and TIME. TimesFM 2.5, from September 2025, is a 200M model with a 16k context.
The licences split by version. Code and weights up to 2.5 are Apache 2.0; the 3.0 weights carry a TimesFM non-commercial licence that bars commercial or production use of downloaded copies. Commercial use of 3.0 runs through Google Cloud, chiefly the AI.FORECAST function in BigQuery. The repo has about 35,500 stars and the last commit was on 13 November 2026.
Is TimesFM worth it?
TimesFM makes a credible forecast the cheapest step in a demand or capacity project, and the licence on its newest version decides how you can use it. A zero-shot model reads a series it has never seen and predicts the next values with no per-series training, which replaces weeks of tuning ARIMA or gradient-boosted models across thousands of products with one batched call.
Version 3.0 closes the gap that kept earlier releases out of serious planning work. Up to 2.5, TimesFM forecast one series from its own history; 3.0 handles several related series at once and accepts covariates such as a promotion calendar. Google's blog shows a mock ice-cream example where a known promotion schedule lifts the forecast by about 20% on promotion days. The benchmark rankings are Google's own, reported as average ranks against Chronos-2, Toto 2.0 and TimesFM 2.5, and the blog gives charts rather than scores.
The catch: the 3.0 weights on Hugging Face are licensed for non-commercial, non-production use. A company that downloads them and runs forecasts for its own stock planning is outside the licence. The commercial route is BigQuery's AI.FORECAST, where 3.0 is still labelled Preview, caps context at 2,048 points against 15,360 when self-hosted, and limits the horizon to 1,024 points. TimesFM 2.5 remains Apache 2.0 and is the default in BigQuery, which is why it still draws about 1.7 million downloads a month against roughly 479,000 for 3.0.
The repo itself is busy and well kept for a research project: releases v2.0.1, v2.0.2 and v3.0.0 between June and August 2026, a test suite added in April, and an MLX backend that the README reports at 666 series a second on an M4 Max. The README also says this open version is not an officially supported Google product, so support is GitHub issues.
Our take: treat 2.5 as the free baseline every forecasting team should beat before building anything bespoke, and treat 3.0 as a Google Cloud product with a research preview on the side.
Where it earns its place
A demand planner at a retailer with 20,000 SKUs
Get a weekly baseline forecast for every product without a modelling team
Batch the sales history through TimesFM 2.5 on one GPU with predict_batch and use the quantiles for safety stock. Apache 2.0, so production use is allowed; cost is the GPU hours.
A data analyst whose sales data already sits in BigQuery
Forecast revenue by region with the marketing calendar as a covariate
Call AI.FORECAST with model set to TimesFM 3.0 and the promotion flags as future covariates, inside one SQL query. Billed at BigQuery rates, with token pricing for 3.0 from 1 December 2026.
An SRE team running a few hundred services
Spot capacity shortfalls a week ahead from CPU and request metrics
Run TimesFM 2.5 over each service's hourly metrics and alert when the 90th percentile forecast crosses capacity. Free to self-host; a small GPU or an Apple silicon laptop covers the batch.
Before you commit
- Setup
- pip install timesfm[torch] (or timesfm[mlx] on Apple silicon) on Python 3.10+, then load google/timesfm-3.0-pytorch or the 2.5 checkpoint from Hugging Face and call predict_batch with NumPy arrays. A first forecast takes minutes.
- Integrations
- Hugging Face checkpoints; PyTorch, Flax and MLX backends; Transformers and PEFT for LoRA fine-tuning; BigQuery ML AI.FORECAST, Connected Sheets and Vertex AI Model Garden on Google Cloud; an AGENTS.md and SKILL.md for coding agents.
- Data and privacy
- Self-hosted inference runs locally and sends nothing to Google. Code and weights up to 2.5 are Apache 2.0; 3.0 weights are under the TimesFM non-commercial licence. BigQuery use falls under the Google Cloud terms, which allow commercial use.
- Cost at scale
- Self-hosting costs only compute; the 200M and 330M models are small enough for one GPU or a laptop. In BigQuery, AI.FORECAST is billed at BigQuery ML rates, and 3.0 moves to token-based pricing from 1 December 2026; per-token prices are not published in the docs we read.
- Limits
- BigQuery caps 3.0 at a 2,048-point context and a 1,024-point horizon (2.5 allows up to 15,360 and 10,000); each series needs at least 3 points. Self-hosted 3.0 truncates contexts over 15,360 points. About 240 open issues and pull requests.
What’s good
- Zero-shot point and quantile forecasts with no per-series training
- Version 3.0 adds multivariate series and past or future covariates
- Versions up to 2.5 are Apache 2.0 for code and weights
- Runs on PyTorch, Flax or MLX, and inside BigQuery and Google Sheets
What’s not
- TimesFM 3.0 weights are non-commercial and non-production only
- Benchmark rankings are self-reported, as average ranks without scores
- BigQuery's 3.0 is Preview with a 2,048-point context
- Not an officially supported Google product; support is GitHub issues
Best for
Analysts and data teams who need decent forecasts across many series quickly, and who can either live with TimesFM 2.5 or already run on BigQuery.
TimesFM alternatives
TimesFM: common questions
- Is TimesFM worth it?
- A forecasting model that tops the leaderboards by Google's own count, with one trap: the 3.0 weights are for research only, so production users either stay on Apache-licensed 2.5 or pay Google through BigQuery. We rate it 4.3 out of 5.
- How much does TimesFM cost?
- Code and weights up to 2.5 are free under Apache 2.0. TimesFM 3.0 weights are free for non-commercial use only; commercial use goes through BigQuery ML, billed at BigQuery rates, with token-based pricing for 3.0 from 1 December 2026.
- What are the downsides of TimesFM?
- TimesFM 3.0 weights are non-commercial and non-production only; benchmark rankings are self-reported, as average ranks without scores; bigQuery's 3.0 is Preview with a 2,048-point context; not an officially supported Google product; support is GitHub issues.
- Who is TimesFM best for?
- Analysts and data teams who need decent forecasts across many series quickly, and who can either live with TimesFM 2.5 or already run on BigQuery.
- How much does TimesFM cost for a team or in production?
- Self-hosting costs only compute; the 200M and 330M models are small enough for one GPU or a laptop. In BigQuery, AI.FORECAST is billed at BigQuery ML rates, and 3.0 moves to token-based pricing from 1 December 2026; per-token prices are not published in the docs we read.
- Is TimesFM safe for company data?
- Self-hosted inference runs locally and sends nothing to Google. Code and weights up to 2.5 are Apache 2.0; 3.0 weights are under the TimesFM non-commercial licence. BigQuery use falls under the Google Cloud terms, which allow commercial use.
- What are the best TimesFM alternatives?
- The closest alternatives we have reviewed are Kronos.
Answers checked Nov 14, 2026.
Make TimesFM? Show this rating on your site
