time series
TimesFM 3.0
Google Research multivariate time-series foundation model; best single zero-shot model (rank 14 of 143) on the GIFT-Eval leaderboard rebuilt from its raw results. Code Apache, weights non-commercial.
Levelmodel
Classtime series
Parameters330M
Contextnot in the notes
Released31 Aug 2026
Opennessopen weights restricted
Runs onsingle datacentre GPU
Statusactive
Benchmark Results
vendor-reportedindependent
fev-benchtime series
Google claims best pretrained FM; no numbers givenvendor-reported
no score in the notes
rank 1
GIFT-Evaltime series
zero-shotindependentno test-data leakage
0.667
00.71.4
Score in MASE (geo-mean, normalised to Seasonal Naive). Lower is better.
Leaderboard rebuilt from the published raw results, 143 entries covering all 97 configurations
zero-shot; rank is average CRPS rankindependentno test-data leakage
0.456
00.51
Score in CRPS (geo-mean, normalised to Seasonal Naive). Lower is better.
rank 14 of 143, Leaderboard rebuilt from the published raw results, 143 entries covering all 97 configurations
Google claims best pretrained FM; no numbers givenvendor-reported
no score in the notes
rank 1
TIMEtime series
Google claims best pretrained FM; no numbers givenvendor-reported
no score in the notes
rank 1
Where this comes from
Every score names who ran it: vendor-reported, maintainer-run or independent. A rank is shown with the population it was ranked in (gap rule 3).
Lineage
successor
TimesFM 2.5 is the parent (successor)high confidencenotesWhere this comes from
Parent and child relations from model cards, Hugging Face base_model fields and the genealogy tables of the research notes. The tree shows two generations each way, 3 models. Confidence and evidence are the notes' own.
Industry Use Cases
R&D Topics
Identity
Where this comes from
Developer and dates from the research notes. Release date is the vendor's public announcement; weights date is when open weights were published.
Architecture and Size
- Architecture
- dense transformer Decoder-only transformer, 32-point patches; alternating within-series causal attention and cross-series attention; masked full-horizon single-pass decoding; native multivariate + covariates
- Total parameters
- 330M sources
- Active parameters
- –
- Checkpoint parameters
- –
- Note
- about 330M; trained on more than 1 trillion time points
- Context length
- not in the notes
- Input
- time series
- Output
- time series
Where this comes from
Parameter counts as the developer states them. Where the checkpoint holds a different total, both are shown (register A16).
Openness and Access
- Openness
- open weights restricted
- Licence
- timesfm-non-commercial-license-v1.0 sources
- Access routes
- open weights download, cloud marketplace
Where this comes from
The licence file governs over registry tags and aggregator labels.
Deployment
- Runs on
- single datacentre GPU
- Footprint
- Production use only via BigQuery; PyPI timesfm 3.0.2 on 9 Sep 2026
- Runtimes
- not in the notes
Where this comes from
Deployment tiers and runtimes as named in the notes. Footprint is the notes' figure, not a measurement by AURION.
Indonesia
- Indonesian support
- not stated
- Regional languages
- not in the notes
- Jakarta hosting
- unknown
- Languages
- not in the notes
Where this comes from
Indonesia fields are carried on every model; unknown is an allowed value.
Links
DeveloperGooglebig tech, USA
Googlebig tech, USAModelTimesFM 3.0Google, time series
TimesFM 3.0Google, time seriesR&D TopicsTime-Series Foundation Models
Who Runs Itnone recordedDeployments of these use cases; none names this model.
Organisations 1
Sources
- thenewstack.io/google-timesfm-3-multivariate-forecasting https://thenewstack.io/google-timesfm-3-multivariate-forecasting/
- TimesFM vs Chronos vs Moirai: The Leader's Weights Ban Production beri.net https://www.beri.net/article/timesfm-vs-chronos-vs-moirai-time-series-foundation-model-licence-bigquery-sagemaker-cost-2026
- GIFT Eval - a Hugging Face Space by Salesforce huggingface.co https://huggingface.co/spaces/Salesforce/GIFT-Eval
- pypi.org/pypi/timesfm/json https://pypi.org/pypi/timesfm/json
- Google TimesFM-3 voegt multivariate forecasting toe datanorth.ai https://datanorth.ai/nl/nieuws/google-lanceert-timesfm-3
- TimesFM-3 – open weights llmreference.com https://www.llmreference.com/model/timesfm-3
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "TimesFM 3.0", https://aurion.semai.id/models/google-timesfm-3-0/, records of October 2026.