InPhroNeSys panther mark
InPhroNeSysActionable Insights
Research snapshot · 8 November 2026
36-month window · 8 Nov 2023 – 8 Nov 2026

The TSFM release observatory

Follow the release pace, inspect what changed, and compare forecasting results within a named benchmark.

Curated milestones in view—
Recent 12 months / preceding 12—
Organizations represented—Within the current filters
Retrospective error improvement—

Release timeline

How quickly is the catalogue growing?

Release catalogue

Read the evidence, definitions and selection rules

What this catalogue counts

  • Selected model generations and substantial public research announcements. It is a curated catalogue, not a complete market census.
  • Model variants released together count once. Moirai 2.0 uses its August weight announcement; its November paper does not add a second event. IBM's March 2026 portfolio counts once.
  • A paper date, public announcement date and first weight upload can differ. The drawer names the date basis. Month-only dates use the 15th for positioning; displayed labels retain month precision.
  • The 36-month window ends on 8 November 2026. TimeGPT and the early TimesFM / Lag-Llama papers appeared in October 2023, before the window. Empty months mean no event in this selection.
  • Recent and preceding 12-month counts use the current end date and active filters, clipped to the selected start date. Incomplete windows are labeled. No exponential growth model has been fitted.

What the performance charts establish

  • Mistral Table 3 reports a selected-model comparison dated 15 October 2026. H2O's model card is a separate publisher snapshot. The selector keeps them separate.
  • GIFT-Eval and TIME relative MASE use geometric-mean errors normalized to seasonal naïve (1.0). fev-bench SQL skill is a different probabilistic metric; higher is better. These scales are never averaged.
  • The date frontier is the best retrospective score among visible checkpoints, evaluated in the selected snapshot. It does not reconstruct the historical SOTA leaderboard.
  • The parameter frontier minimizes parameter count and forecast error (or maximizes SQL skill). Parameter count is not latency, total cost or an engineering deployment benchmark.
  • No synthetic scores, Monash/GIFT composite, 40% historical improvement claim, independent rerun or significance claim is included. Missing values stay missing. The current subset does not prove superiority to locally fitted ETS or ARIMA.

Practical next step: compare SNAIVE, ETS / ARIMA and a suitable TSFM on identical rolling forecast origins and purchasing horizons. Check bias, prediction intervals, service / inventory cost and the exact checkpoint license. FPP3 cross-validation guide.