Google's TimesFM-3 tops forecasting — but weights go non-commercial
The 330M open model wins GIFT-Eval, FEV-Bench and TIME zero-shot and finally forecasts multivariate — plus Runway's Solaris renders live UIs with no code.
Copy markdown
Zero-shot, and now multivariate
TimesFM-3 is a 330M-parameter foundation model that forecasts multiple correlated series at once — with past and future covariates like promotions or weather — in a single non-autoregressive pass, no training run required. Point it at demand planning, capacity, or observability data and get predictions cold.
It's #1 on the boards it entered
It ranks first on GIFT-Eval, FEV-Bench, and the TIME-leaderboard across both point and probabilistic forecasting, beating Amazon's Chronos-2, Salesforce's Moirai 2.0, and Datadog's Toto 2.0 — and it still wins in plain univariate mode against its own predecessor.
The license quietly flipped to non-commercial
Here's the catch builders will miss: the TimesFM-3 weights ship under a new non-commercial license, a step back from TimesFM-2.5's Apache 2.0. The code stays Apache 2.0, but production use of the pretrained weights isn't allowed — Google's paid path is BigQuery's AI.FORECAST. Prototype freely on Hugging Face; check the license before you ship.
Elsewhere: Runway's Solaris renders UIs as live video
Runway unveiled Solaris, an 'interface world model' on Gen-4.5 that generates interactive app UIs frame-by-frame as video — no code, no fixed screens, just a responsive scene that reacts to clicks and drags. It beat coded interfaces 61% to 24% on instruction-following in Runway's study, but it's early-access-by-form only, with no API, pricing, or public date yet — a glimpse, not a build.