TimesFM 2.5 vs 3.0: Licensing and Commercial Use Explained
TimesFM 2.5 is Apache 2.0 licensed, but TimesFM 3.0's self-hosted weights are non-commercial. Here's what that means for production use.
What is TimesFM, and why does the license matter?
TimesFM is Google Research’s foundation model for time series forecasting, a decoder-only transformer trained on time series data instead of text. It forecasts zero-shot, meaning it can produce a forecast for a new series (daily revenue, server load, ridership) without being trained on that specific data first. The license matters because it determines whether you can legally run the model in a commercial product, and Google changed the terms between versions 2.5 and 3.0 in a way that directly affects teams trying to build on it.
TL;DR
- TimesFM 2.5 ships under the Apache 2.0 license, which permits commercial use of the self-hosted weights with no special agreement needed.
- TimesFM 3.0’s self-hosted weights are non-commercial only, so running the downloaded model in a revenue-generating product isn’t permitted under that license.
- Commercial access to 3.0 is routed through Google Cloud and BigQuery rather than through self-hosting the open weights.
- In testing against real New York transit ridership data, both versions cut forecasting error by about 25% compared to a seasonal baseline with no training on the data.
- On a simulated revenue forecast, TimesFM 3.0 landed within 2% of actual results at a short horizon and stayed within 6% a full quarter out.
- The two versions can give opposite answers on what-if planning questions (like a marketing budget increase), which matters more for production decisions than raw accuracy.
- Version 3.0 added the ability to condition forecasts on known future inputs, like holidays or promotions, directly inside the model rather than through a separate regression step.
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
What changed between TimesFM 2.5 and 3.0?
The core architecture stayed the same: a transformer that chops historical data into patches (TimesFM uses patches of 32 time steps) and treats each patch like a token, predicting the next chunk of the chart much like a language model predicts the next word. TimesFM 3.0 was reportedly pre-trained on more than a trillion time points across real and synthetic datasets, per Google.
The functional change in 3.0 is how it handles information you already know about the future. In version 2.5, if you wanted to account for an upcoming holiday or promotion, you had to run a separate regression step outside the model first, then let TimesFM forecast whatever the regression didn’t explain. Version 3.0 folds that mechanism into the model itself. You can feed it multiple related series at once, along with past covariates and future inputs like a planned price change or a holiday calendar, and the transformer looks ahead at those inputs before generating a forecast. It also produces the entire forecast horizon in one pass instead of building it step by step, which makes it noticeably faster than 2.5 in practice.
Google has also claimed 3.0 ranks first among foundation models across three public forecasting benchmarks, though independent testing on real-world data (outside the model’s training window) is where the practical differences between versions actually show up.
How do the licenses differ in practice?
This is the part that matters most if you’re deciding what to build with.
TimesFM 2.5 is released under Apache 2.0, a permissive open-source license. That means you can download the weights, self-host the model, and use it in a commercial product without asking Google for permission or paying a licensing fee.
TimesFM 3.0 changes that arrangement. The self-hosted weights for 3.0 are licensed for non-commercial use only. If you want to use the newer model commercially, Google routes that access through Google Cloud and BigQuery instead of letting you freely self-host the open weights. In effect, the faster and more capable model comes with a narrower path to commercial deployment, while the older model remains the one you can drop into a production pipeline on your own infrastructure without a cloud dependency.
For teams evaluating which version to build on, this turns into a tradeoff between capability and control. TimesFM 3.0 is faster and can natively incorporate future known events like holidays and promotions. TimesFM 2.5 is slower and needs an external regression step for the same thing, but it’s unambiguously free to self-host for commercial purposes.
Is TimesFM reliable enough to use in production?
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
For straightforward forecasting, testing against real-world data suggests yes. On daily ridership data from New York’s subway, buses, and commuter rail (data cut off before either model’s training window), both versions reduced forecasting error by roughly 25% compared to a simple seasonal baseline, with zero training on the specific dataset. Feeding in a holiday calendar as a future input improved both models further, and both correctly adjusted their Labor Day forecast downward once given that context, something neither model did from historical patterns alone.
In a separate simulation using nearly four years of daily trials and revenue data for a hypothetical software business, TimesFM 3.0 forecast two months out and landed within 2% of actual revenue. Run again a full quarter later, it stayed within 6%, with both models’ predicted uncertainty ranges holding up.
The reliability story gets more complicated once you move from “continue this existing pattern” to “tell me what happens if we change something.” That’s where the two versions can diverge sharply, and where the choice of model (and its license) becomes a business risk question, not just a technical one.
Where does zero-shot forecasting break down?
The failure mode shows up on what-if planning questions rather than on routine forecasting. In a test comparing two marketing proposals (a deeper Black Friday promotion versus doubling the January marketing budget), TimesFM 3.0 barely reacted to the doubled January budget, while TimesFM 2.5 predicted a large revenue jump from the same change. An AI agent given both forecasts trusted version 3.0 and recommended the wrong proposal, missing the plan that actually produced more revenue in the simulation’s ground truth. The true outcome sat between the two models’ predictions, and testing four different ways of encoding the marketing spend didn’t fix the underreaction in 3.0.
This points to a general limitation: a zero-shot model can only extrapolate levers that have actually moved somewhere in the training history it’s drawing on. If a variable like price has never changed in your own historical data, no foundation model, zero-shot or classically trained, can reliably infer your customers’ price sensitivity. Large out-of-distribution moves (a budget or promotion that goes well beyond anything in the historical range) are exactly where version 2.5 and 3.0 can produce conflicting answers, which argues for sanity-checking any what-if forecast against past campaign results rather than trusting a single model’s output. It’s also worth backtesting the model’s uncertainty ranges on your own historical data before relying on them, since confidence intervals that look fine on one dataset can run narrow on another.
Frequently Asked Questions
Can I use TimesFM 2.5 commercially for free?
Yes. TimesFM 2.5 is released under the Apache 2.0 license, which allows commercial use of the self-hosted weights without a separate commercial agreement.
Can I self-host TimesFM 3.0 for a commercial product?
The self-hosted weights for TimesFM 3.0 are licensed for non-commercial use. Commercial use of 3.0 is intended to go through Google Cloud and BigQuery rather than through freely self-hosting the downloaded weights.
Is TimesFM 3.0 more accurate than 2.5?
Not necessarily. In testing, the two versions performed similarly on continuing existing patterns, with 3.0 running several times faster. On what-if planning scenarios involving large changes to a variable, the two versions gave notably different, sometimes opposite, predictions.
Does TimesFM need to be trained on my data before I use it?
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
No. TimesFM is designed to forecast zero-shot, meaning it can generate forecasts for a new time series without being fine-tuned on that series first. It works by recognizing general patterns like seasonality, trends, and holiday effects rather than needing domain-specific training.
What’s the main risk of trusting TimesFM for business planning?
The main risk is relying on it for decisions involving changes your historical data has never seen, like a price increase that’s never happened before or a budget increase far outside past spending levels. In those cases, forecasts from different model versions can diverge significantly, so it’s safer to validate with controlled experiments or backtests rather than trusting a single zero-shot forecast.
