Products/AI/ML - Foundation Model / Time Series Forecasting/timesfm

timesfm

A pretrained time-series foundation model developed by Google Research for time-series forecasting

AI/ML - Foundation Model / Time Series ForecastingMountain View, USAPrivate company (part of Alphabet Inc.)thousands peopleFounded 1998Raised Not applicable200M parameters (v2.5)Up to 16k context length supportContinuous quantile forecast up to 1k horizon with optional 30M quantile headFlax version for faster inferenceCovariate support via XRegFine-tuning with LoRA via HuggingFace Transformers + PEFTUnit tests for core layers, configs, and utilitiesDecoder-only architecturePublished at ICML 2024
Visit timesfm →
timesfm

Our Take

This is Google's time-series foundation model — 200M params, 16k context window, decoder-only architecture, free to use. It's got 17.5k stars which is real traction, and it's already baked into BigQuery ML and Google Sheets so enterprises can forecast without building from scratch. The ICML 2024 paper signals Google is actually serious about this space, not just experimenting. TimesFM 2.5 is the move if you're already living in Google's ecosystem and need time series forecasting without the headache.

Decoder-only foundation model for time-series forecasting that can analyze historical time series data and predict future values

Problem It Solves
Time series forecasting - enables users to forecast future values from historical time series data without building models from scratch
Target Customer
Data scientists, ML engineers, enterprises needing time series forecasting capabilities
Use Cases
Enterprise time-series forecasting, SQL-based forecasting via BigQuery ML, Spreadsheet forecasting via Google Sheets, Agentic calling via Vertex Model Garden, Fine-tuning for domain-specific forecasting
Differentiator
Google Research-developed pretrained foundation model with 16k context length, supports continuous quantile forecasting with optional quantile head
Why Now
Foundation models for time series is an emerging field; TimesFM 2.5 brings significant improvements over previous versions including larger context window and efficient model size
Traction
Customers Mentioned: BigQuery ML (Google), Google Sheets (Google), Vertex Model Garden (Google) · Notable Metrics: 17.5k stars, 1.7k forks, 320 commits, 137 issues, 58 pull requests · Press Mentions: ICML 2024 paper

Key Facts

Category
AI/ML - Foundation Model / Time Series Forecasting
Location
Mountain View, USA
Founded
1998
Team Size
thousands people
Stage
Private company (part of Alphabet Inc.)
Raised
Not applicable
Pricing
Free
Discovered via
github-trending

The people behind timesfm

@

@borealBytes

profile
@

@darkpowerxo

profile
@

@kashif

profile

Links

WebsiteGitHubSource: github-trending

Want products like this in your inbox every morning?

Five products. Every morning. Written by someone who actually cares whether they're good or not. Free forever, unsubscribe whenever.