March 30, 2026
Eggs, inflation, and 16k vibes
Google's 200M-parameter time-series foundation model with 16k context
Smaller model, bigger memory — fans cheer, skeptics ask “predict eggs AND inflation?”
TLDR: Google’s TimesFM 2.5 shrinks to 200M parameters, stretches memory to 16k, and adds uncertainty forecasts, but the open repo isn’t officially supported. Commenters split between curiosity and skepticism, asking for real-world proof, explainability, and apps—while others point to Prophet and Nixtla and say “call me when it ships results.”
Google just dropped TimesFM 2.5, a leaner 200M‑parameter time‑series model with a jumbo 16k memory window and an optional add‑on that gives uncertainty ranges. There’s even an official BigQuery tie‑in, though the open repo is “not officially supported.” The code is on GitHub, with a faster Flax version teased and covariate support restored via XReg. So why’s the crowd buzzing? Because this “predict anything” promise is either magic or marketing.
The hottest debate: one skeptic asks how a single model can forecast both “egg prices in Italy” and “global inflation,” and wonders where the explanations are. Translation: trust issues. Meanwhile, curious onlookers want competitions and tutorials (“I’ll try it, time series is hard”), while pragmatists say they’ll stick with old reliables like Nixtla and Prophet until TimesFM proves it in the wild. Another jab: “Has anyone built anything on it?” The subtext is clear—show us real users, not just parameter counts.
The vibe is classic: Google ships lab wizardry; the internet asks for receipts. Some love the long memory and the “quantile” uncertainty, likening it to weather‑style confidence bands. Others want explainability, benchmarks, and production wins. Until then, TimesFM 2.5 is a shiny crystal ball that the crowd can’t decide is genius or just a neat snow globe.
Key Points
- •Google Research released TimesFM 2.5, a 200M-parameter decoder-only time-series foundation model.
- •TimesFM 2.5 supports a 16k context length and optional continuous quantile forecasting up to a 1,000-step horizon via a 30M quantile head.
- •The release removes the frequency indicator and introduces new forecasting flags; the inference API has been upgraded.
- •TimesFM is available as an official Google product in BigQuery; the open-source release is not officially supported by Google.
- •Installation supports PyTorch and Flax (JAX) backends, with provided example code and guidance for different hardware (CPU, GPU, TPU, Apple Silicon).