Google Releases TimesFM-3: A 330M Multivariate Time Series Foundation Model
Balance- · reddit · 2026-09-04
Google Research released TimesFM-3, a 330M-parameter zero-shot time series foundation model whose headline change is native multivariate forecasting — multiple targets, past-only covariates, and past-future covariates (holidays, planned promotions) without fine-tuning.
Architecture
- Decoder-only Transformer, 20 layers, dim 1280, 16 heads, 32 timesteps patched per token
- Alternating attention per layer: causal across time within a series, full attention across series at each timestep
- One-forward-pass forecasting: masked placeholder tokens for the whole horizon are filled simultaneously; outputs 9 quantiles (P10–P90) per target per step
Pretraining & results
- Trained on GiftEvalPretrain, Wikipedia pageviews (to Nov 2023), Google Trends (to end 2022) plus synthetic data — over 1 trillion time points
- Best average rank on Gift-Eval and FEV-Bench vs Chronos-2, Toto 2.0, TimesFM-2.5; univariate-only mode already matches or beats baselines
Caveat: weights ship under the non-commercial TimesFM license — not production-ready. PyTorch weights are on Hugging Face and GitHub; BigQuery integration coming later.
More from Models
- Opus 5.5 tops OpenRouter spending in 20 of 29 task categories — mmurph · 2026-10-21
- Meta team builds open-source 30B LLM from scratch in months, hails 'second era' — TimDarcet · 2026-10-21
- DFlash drafter for Muse Glimmer-30B: 5 layers, blockwise 16-token spec decoding — TimDarcet · 2026-10-21
- EvoPress quantization shrinks 30B model from 56GB to 19GB at near-bf16 accuracy — TimDarcet · 2026-10-21
- Radar chart: Muse Glimmer-30B benchmarked against Gemma4-31B and Qwen3.6-27B — TimDarcet · 2026-10-21
- Privacy evals: Muse Glimmer-30B leaks less than Qwen3.6-27B on CIMemories — TimDarcet · 2026-10-21