Article
Google TimesFM-3: The best forecasting model you can't use in production
Why this matters: Most people think foundation models are about chatbots and coding assistants. But there's a quieter, arguably more commercially important race happening in time-series forecasting — the kind of math that decides how much inventory a retailer orders, how much energy a grid needs to buy, or how a financial model moves with interest rates. Google just released the strongest entry yet. With a catch.
What TimesFM-3 is
TimesFM-3 is a 330-million-parameter time-series foundation model from Google Research that forecasts multiple related series simultaneously in a single forward pass — no task-specific fine-tuning required [1]. It was pretrained on more than one trillion time points of real and synthetic data [2].
Every prior version of TimesFM (through v2.5, released September 2025) was univariate: one series, its own history, nothing else [3]. That's a problem because almost no real forecasting problem is shaped that way. Store-level demand across a retail chain, or interest-rate-sensitive line items across a company's financial model — those are inherently multivariate, with series that move together [4].
How it works (the architecture, briefly)
The backbone is a decoder-only transformer. Time points are grouped into patches of 32 steps, then normalized per series so wildly different scales don't dominate [1]. The model accepts three input types zero-shot:
- Multiple targets — forecast jointly with point and quantile outputs for each
- Past covariates — known only historically (e.g. past foot traffic)
- Past-future covariates — whose future values are known (e.g. a promotion calendar, holiday schedule)
Tokens then enter a 2D grid and pass through two alternating attention mechanisms [2]:
- Causal temporal attention (horizontal) — strictly causal, confined to earlier tokens inside the same series. Blocks leakage.
- Full variate attention (vertical) — at a given time step, a token reads every other series at that step, learning cross-series correlations.
This alternating design is the key innovation: it lets the model learn relationships between series without leaking future information.
One forward pass instead of many
Earlier TimesFM versions decoded one patch at a time — adding latency, compute cost, and compounding error. TimesFM-3 uses Contiguous Patch Masking (introduced with TiRex): masked placeholder tokens are appended for the whole horizon, and the alternating attention layers fill every masked patch simultaneously [2]. Each target receives 9 quantiles (10th through 90th percentile) at every horizon step — not just a point forecast, but a distribution.
Benchmarks — it genuinely leads
Google evaluated on three major time-series foundation model benchmarks [1][3]:
| Benchmark | TimesFM-3 result | |-----------|-----------------| | fev-bench | Rank #1 overall across 100 real-world forecasting tasks | | TIME Benchmark | Rank #1 overall across 50 domain datasets, 98 evaluation tasks | | GIFT-Eval | Rank #1 among all foundation models (point and probabilistic metrics) |
It outperforms Amazon's Chronos, Salesforce's Moirai 2.0 family, and its own predecessor TimesFM-2.5 [4].
The catch — you can't ship it
Here's the part that matters if you're building a product:
- Code: Apache-2.0 license — you can use, modify, and distribute it freely [5].
- Weights: timesfm-non-commercial-license-v1.0 — restricted to non-commercial, non-production use [5].
You can benchmark TimesFM-3 today. You cannot ship it behind a production forecast API unless you negotiate separate licensing with Google. The prior version, TimesFM-2.5 (200M parameters, Apache-2.0 weights), remains the safe option for commercial deployment [3].
Competitive landscape
TimesFM-3 lands in a smaller, less-hyped corner of the foundation-model race than chatbots, but it's a real one [4]:
- Amazon Chronos — open forecasting foundation model
- Salesforce Moirai — open forecasting model, now at v2.0
- Nixtla — specialized forecasting infrastructure for retail/supply-chain
The VC read (from Value Add VC): "The diligence question for any forecasting-infrastructure startup right now is: what do you do better than a free, Apache-licensed model from Google, Amazon or Salesforce? Data integration and vertical tuning are the only durable answers left — raw model quality stopped being a moat in this category the moment three hyperscalers all shipped competent open forecasting models in the same year." [4]
What to watch next
The number worth tracking is adoption inside Google Cloud's own Vertex AI forecasting products — that would signal whether TimesFM-3 is primarily a research showcase or an actual commercial wedge against Amazon and Salesforce's competing models [4]. If Google keeps the weights non-commercial but bakes the model into BigQuery ML and Vertex AI (where they already offer TimesFM via SQL queries [6]), the competitive pressure on standalone forecasting vendors intensifies without Google ever needing to sell the model directly.
Sources
- Google Research blog — "TimesFM-3: A zero-shot foundation model for multivariate forecasting" (Aug 31, 2026)
https://research.google/blog/timesfm-3-a-zero-shot-foundation-model-for-multivariate-forecasting/
- MarkTechPost — "Google AI Releases TimesFM-3: A 330M Parameter Zero-Shot Foundation Model For Multivariate Time Series Forecasting" (Aug 31, 2026)
- GitHub — google-research/timesfm — README, benchmark results, licensing details
- Value Add VC — "Google Quietly Ships a Foundation Model for Forecasting" (Aug 31, 2026)
https://valueaddvc.com/pulse/google-timesfm-3-forecasting-model-launch-2026
- TimesFM 3.0 license — timesfm-non-commercial-license-v1.0 (weights); Apache-2.0 (code)
https://github.com/google-research/timesfm#license-notice-for-pretrained-weights
- Google Cloud BigQuery ML docs — TimesFM model documentation
- Google Research on X — TimesFM-3 announcement (Aug 31, 2026)