TimesFM-3

TimesFM-3 is Google Research’s latest time-series foundation model, a roughly 330M-parameter forecasting model designed for zero-shot use. The major change from earlier TimesFM generations is native multivariate forecasting: it can forecast multiple target series together and use both past-only and known-future covariates without per-task fine-tuning. It produces point and quantile forecasts and can also operate in univariate mode. ¹²

This makes TimesFM-3 materially more useful for real forecasting problems such as retail demand with promotions, operational metrics with cross-series relationships, finance, manufacturing and other settings where future-known variables or correlated series matter. The model is distributed for local/programmatic use through Google’s repository, Hugging Face weights and the TimesFM Python package. ¹³²⁴

Benchmark performance is genuinely strong but needs careful framing. Google reports leading average ranks among pretrained foundation models on GIFT-Eval, FEV-Bench and TIME. The live GIFT-Eval leaderboard corroborates a strong submission but currently places TimesFM-3 around tenth overall in the broader field, behind several agent/wrapper or other submitted systems. AiToolMap therefore treats TimesFM-3 as a top-tier pretrained foundation model, not as an uncontested #1 forecasting system across every submission category. ¹⁵⁶

There is no conventional SaaS price. TimesFM 3.0 code/package is available publicly and the source code is Apache-2.0, but the pretrained 3.0 weights use the separate TimesFM Non-Commercial License v1.0. The official repository and model card explicitly restrict those weights to non-commercial, non-production use. ³²⁴

For research and evaluation, that means the entry price is effectively zero: users can download the package and weights and run inference on their own compute. For a company looking for a production forecasting model, however, “free” is misleading because the default weights cannot legally be deployed for commercial/production use. Google also says a BigQuery integration is coming, but as of this review it is not yet the current TimesFM-3 access path. ¹²

TimesFM-3 has an unusually transparent evidence surface for a model this new. Google publishes the architecture/capability summary, source code, package, model weights, training-data categories and an explicit model license. The Hugging Face card identifies the 3.0 license clearly and currently shows no inference-provider deployment, which makes the access boundary easy to verify rather than infer. ¹³²⁴

The public specialist benchmark evidence is useful because it prevents over-reading Google’s own charts. The current GIFT-Eval raw/leaderboard surface shows TimesFM-3 performing strongly, but broader rankings include systems ahead of it. This is exactly the kind of context that matters for a technical model review: vendor claims about leadership are credible within their comparison class but are not a universal performance guarantee. ⁵⁶

The fixed AiToolMap 50-source general panel produced no qualifying exact-model evidence, which is unsurprising for a specialist forecasting model only released days ago. Forty-five families returned no relevant result and five were inaccessible. That lack of user/product-market evidence keeps rating confidence LOW even though technical benchmark evidence is substantially better than for most newly launched tools. ¹⁵

TimesFM-3 is best for forecasting teams that can run Python models themselves and want a strong zero-shot baseline for multivariate problems before investing in task-specific training. Researchers comparing foundation models also benefit from the public code, benchmark surfaces and transparent model card. ³²⁴

It is not yet a straightforward choice for a production SaaS, enterprise planning system or commercial analytics pipeline because the default 3.0 weights are explicitly non-production/non-commercial. Teams that need a commercially deployable Google forecasting service should wait for a suitable licensed surface or evaluate alternatives with production-compatible terms rather than assuming the Apache code license also covers the weights. ³²