TimesFM 2.5

TimesFM 2.5 is the third open checkpoint of Google Research's foundation model for time-series forecasting, announced in the project repository on 15 September 2025. Like the rest of the line it forecasts zero-shot: it is shown a history of measurements and returns a horizon, without being trained on the user's data. The generation is unusual in that it is smaller than its predecessor and better at the same time — 200 million parameters instead of the 500 million of TimesFM 2.0, with a context window grown from 2,048 points to 16,384. Google also removed the frequency indicator that earlier versions demanded from the user, so a series no longer has to be declared as hourly, daily or monthly before it can be forecast. Architecturally it is a decoder-only transformer over patches: 20 layers of dimension 1,280, 16 attention heads, history read in patches of 32 points, a base horizon of 128. On top of that sits an optional quantile head of roughly 30 million parameters, which extends probabilistic forecasting to a horizon of 1,024 points and returns nine quantiles from 0.1 to 0.9 instead of a single number — which is why the weights file counts 231 million parameters rather than the 200 million in the model's name. Training data is the GiftEvalPretrain collection, Wikipedia pageviews to November 2023, the most frequent Google Trends queries to the end of 2022, and synthetic and augmented series. The model kept growing after launch: a Flax version for faster inference, covariate support returned through XReg in October 2025, a LoRA fine-tuning example on Hugging Face Transformers with PEFT, and an MLX backend that runs on Apple silicon with no PyTorch at all. What matters most for the reader, though, is the licence. The 2.5 weights are Apache 2.0, exactly as in versions 1.0 and 2.0 — free to deploy in production, commercially and redistributable. The generation that followed, TimesFM 3.0, breaks with that: its weights carry a non-commercial licence and may not be redistributed at all. For anyone who needs a forecasting model in a commercial product without a separate agreement with Google, 2.5 is the newest usable checkpoint, and this is visible in the numbers — it is downloaded twice as often as its successor.

▸Videos

No videos yet.

▸Reviews