--- title: Google’s new forecasting model beats everyone. You can’t... description: Google released TimesFM-3, a 330-million-parameter zero-shot multivariate time-series forecasting model trained on over a trillion data points, which... canonical: https://daily.dev/posts/google-s-new-forecasting-model-beats-everyone-you-can-t-use-it-at-work-yet--cbo00r1uz twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Google’s new forecasting model beats everyone. You can’t use it at work (yet). | daily.dev og:description: Google released TimesFM-3, a 330-million-parameter zero-shot multivariate time-series forecasting model trained on over a trillion data points, which... og:url: https://daily.dev/posts/google-s-new-forecasting-model-beats-everyone-you-can-t-use-it-at-work-yet--cbo00r1uz og:image: https://api.daily.dev/og/posts/cBo00r1UZ.png og:image:alt: Google’s new forecasting model beats everyone. You can’t use it at work (yet). og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Google’s new forecasting model beats everyone. You can’t use it at work (yet). **[The New Stack](https://daily.dev/sources/newstack)** · 4 min read · 0 upvotes · 0 comments ## Summary Google released TimesFM-3, a 330-million-parameter zero-shot multivariate time-series forecasting model trained on over a trillion data points, which outperforms rivals like Amazon's Chronos-2, Salesforce's Moirai 2.0, and Datadog's Toto 2.0 on benchmarks such as Gift-Eval and FEV-Bench. The model uses a decoder-only transformer with alternating causal and cross-series attention layers, and generates the entire forecast horizon in a single forward pass instead of patch-by-patch decoding. Unlike its predecessor TimesFM-2.5, which shipped under Apache 2.0, the new pretrained weights are released under a non-commercial license restricting production use, while Google plans to use the model to power its paid BigQuery ML AI.FORECAST feature. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What is Google's TimesFM-3 model and how does it compare to previous forecasting models? TimesFM-3 is a 330-million-parameter time-series forecasting model trained on over a trillion real-world and synthetic data points, released by Google. It is natively pre-trained for zero-shot multivariate forecasting, meaning it can ingest multiple related time series plus external features simultaneously. It outperforms rivals including Amazon's Chronos-2, Salesforce's Moirai 2.0, and Datadog's Toto 2.0 on benchmarks like Gift-Eval and FEV-Bench, and even surpasses its own predecessor TimesFM-2.5, which was state-of-the-art when it launched in September 2025. _Teams evaluating forecasting models can track releases like TimesFM-3 and licensing shifts on daily.dev._ ### Can I use Google's TimesFM-3 model in a commercial production application? No, commercial or production use of the default pretrained weights is not permitted. TimesFM-3 is distributed on Hugging Face under a separate non-commercial license (timesfm-non-commercial-license-v1.0) restricting it to non-commercial use, a departure from its predecessor TimesFM-2.5 and competitors like Toto 2.0 and Chronos-2, which ship under Apache 2.0. The source code remains Apache-licensed, but the weights themselves carry the restriction. _Developers weighing licensing terms before adopting a model can follow these shifts on daily.dev._ ### How does TimesFM-3's decoding process differ from earlier forecasting models? TimesFM-3 generates an entire forecast horizon in a single forward pass rather than one patch at a time. Earlier versions decoded forecasts patch by patch, which added latency and compounded errors along the sequence. The new model appends masked placeholder tokens for the whole forecast horizon and fills them all in at once, using alternating causal (within-series) and cross-series attention layers over 32-point patches. _Engineers comparing forecasting architectures can follow model design changes like this on daily.dev._ --- Tags: [#machine-learning](https://daily.dev/tags/machine-learning), [#open-source](https://daily.dev/tags/open-source), [#google](https://daily.dev/tags/google), [#time-series-forecasting](https://daily.dev/tags/time-series-forecasting) [View this post on daily.dev](https://daily.dev/posts/google-s-new-forecasting-model-beats-everyone-you-can-t-use-it-at-work-yet--cbo00r1uz) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Google’s new forecasting model beats everyone. You can’t use it at work (yet).","url":"https://daily.dev/posts/google-s-new-forecasting-model-beats-everyone-you-can-t-use-it-at-work-yet--cbo00r1uz","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/google-s-new-forecasting-model-beats-everyone-you-can-t-use-it-at-work-yet--cbo00r1uz"},"datePublished":"2026-08-31T19:47:14.345Z","dateModified":"2026-08-31T20:18:15.394Z","description":"Google released TimesFM-3, a 330-million-parameter zero-shot multivariate time-series forecasting model trained on over a trillion data points, which...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/cce69c06bfe543aa26313796121b5c8f?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/cce69c06bfe543aa26313796121b5c8f?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"The New Stack","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"The New Stack","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/newstack","url":"https://daily.dev/sources/newstack"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/google-s-new-forecasting-model-beats-everyone-you-can-t-use-it-at-work-yet--cbo00r1uz","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"machine-learning,open-source,google,time-series-forecasting","timeRequired":"PT4M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"The New Stack","item":"https://daily.dev/sources/newstack"},{"@type":"ListItem","position":3,"name":"Google’s new forecasting model beats everyone. You can’t use it at work (yet)."}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/google-s-new-forecasting-model-beats-everyone-you-can-t-use-it-at-work-yet--cbo00r1uz#faq","mainEntity":[{"@type":"Question","name":"What is Google's TimesFM-3 model and how does it compare to previous forecasting models?","acceptedAnswer":{"@type":"Answer","text":"TimesFM-3 is a 330-million-parameter time-series forecasting model trained on over a trillion real-world and synthetic data points, released by Google. It is natively pre-trained for zero-shot multivariate forecasting, meaning it can ingest multiple related time series plus external features simultaneously. It outperforms rivals including Amazon's Chronos-2, Salesforce's Moirai 2.0, and Datadog's Toto 2.0 on benchmarks like Gift-Eval and FEV-Bench, and even surpasses its own predecessor TimesFM-2.5, which was state-of-the-art when it launched in September 2025. Teams evaluating forecasting models can track releases like TimesFM-3 and licensing shifts on daily.dev."}},{"@type":"Question","name":"Can I use Google's TimesFM-3 model in a commercial production application?","acceptedAnswer":{"@type":"Answer","text":"No, commercial or production use of the default pretrained weights is not permitted. TimesFM-3 is distributed on Hugging Face under a separate non-commercial license (timesfm-non-commercial-license-v1.0) restricting it to non-commercial use, a departure from its predecessor TimesFM-2.5 and competitors like Toto 2.0 and Chronos-2, which ship under Apache 2.0. The source code remains Apache-licensed, but the weights themselves carry the restriction. Developers weighing licensing terms before adopting a model can follow these shifts on daily.dev."}},{"@type":"Question","name":"How does TimesFM-3's decoding process differ from earlier forecasting models?","acceptedAnswer":{"@type":"Answer","text":"TimesFM-3 generates an entire forecast horizon in a single forward pass rather than one patch at a time. Earlier versions decoded forecasts patch by patch, which added latency and compounded errors along the sequence. The new model appends masked placeholder tokens for the whole forecast horizon and fills them all in at once, using alternating causal (within-series) and cross-series attention layers over 32-point patches. Engineers comparing forecasting architectures can follow model design changes like this on daily.dev."}}]} ```