Google Research released TimesFM 3.0 on Hugging Face on September 1: a 330M-parameter time series foundation model that ranks first on Gift-Eval, FEV-Bench, and the Time leaderboard — all three major benchmarks, zero-shot. Time series is one of the few foundation-model tracks with real enterprise demand beyond LLMs, and Google keeps winning it with a model smaller than GPT-2.
What’s new in 3.0
Native multivariate forecasting, for the first time. Earlier versions predicted one series at a time; 3.0 handles multiple related series together, plus covariates — historical features and known future events like promotions or weather. Under the hood: a 20-layer transformer alternating temporal attention with variate attention, trained on over 1 trillion time points. A Contiguous Patch Masking trick generates the entire forecast horizon in one forward pass, no iterative loop.
How to use it
Weights sit at google/timesfm-3.0-pytorch, and pip install timesfm gets you inference in a few lines. Feed it retail demand, financial series, or server metrics — no fine-tuning needed. BigQuery integration is coming.
One catch: the weights ship under a non-commercial license. Research and evaluation yes, production no. That’s open weights, not open source — read the fine print first.
You Might Also Like
- Google Timesfm Turns Bigquery Into a Forecasting Engine 200m Parameters Zero Training 16k Context
- Google Turboquant Squeezes llm Cache to 3 Bits 6x Less Memory 8x Faster Zero Accuracy Loss
- Ggml Llama cpp Joins Hugging Face and Honestly it was Only a Matter of Time
- Google Pomelli Photoshoot Just Dropped and Product Photographers Should be Paying Attention
- Hume ai Open Sources Tada an llm Based tts With Zero Hallucinations and 0 09 rtf
Top AI Product 
Leave a comment