AI Weekly Malaysia

IBM Research published a Granite Timeseries Ensemble recipe on Hugging Face that combines four Granite time-series models behind one interface and returns point plus quantile forecasts, with no retraining and no fitted ensemble weights. You use it via `pip install "granite-tsfm>=0.3.11"` and `QuantileEnsembleForecaster.from_pretrained("ibm-granite/granite-timeseries-ensemble-r1")`, with the demo running context_length 1024 and prediction_length 24 on the ETTh1 dataset. Two variants are offered: a default Probabilistic Ensemble using uniform probability pooling, and an adaptive IQR Ensemble; the recipe stores only the ensemble configuration, so the four member checkpoints are downloaded at runtime on first run.

Why: If you're shipping forecasting into a product, this gives you a permissively licensed default that removes the 'which model do I pick' step for smooth industrial signals, noisy retail demand, and long seasonal horizons alike. The practical catch is stated in the post itself: because the recipe holds configuration rather than merged weights, your first invocation pulls four checkpoints, so budget for that in cold-start, serverless, or CI environments rather than assuming a single small download. The text contains no Malaysian or Southeast Asian angle, so treat this purely as a tooling decision.

Mistral released Timestral, its first time-series foundation model family, in four sizes from 5M to 209M parameters, with the 209M model claiming first place on GIFT-Eval, TIME and fev-bench while using two-thirds the parameters of the prior state of the art. It produces probabilistic, multivariate forecasts, supports past-only and known-future covariates, and is pitched at CPU inference and zero-shot use, with domain fine-tuning through Mistral Forge. The results have only been submitted to the GIFT-Eval and TIME leaderboards, so the ranking claims are not yet independently verified.

Why: If you maintain forecasting for capacity planning, sensor telemetry or observability metrics, this is a 209M-parameter model you could plausibly run on CPU zero-shot instead of training PatchTST-style models from scratch per dataset — but the benchmark wins are self-reported and pending leaderboard verification, so benchmark it against your own series and your existing ARIMA baseline before replacing anything. Nothing in the announcement addresses Malaysia or SEA deployment, pricing, licensing or data residency, so treat those as open questions rather than settled.

NVIDIA released Kumo Forecast (formerly NV-Tesseract Forecasting), a multivariate time-series forecaster built on a MOMENT-1-large encoder of roughly 340M parameters plus a trainable forecasting head, under Apache 2.0. Its optional DARR mode blends direct forecasts with nearest-neighbor patterns retrieved from a separate historical DataFrame, and it adds horizon-specific interpretability without changing the underlying forecaster.

Why: If you build forecasting into a SaaS product or internal dashboard, this gives you an Apache-2.0 model with a DataFrame-first API, configurable forecast horizons, and a retrieval path you can tune against direct predictions. Treat the vendor blog as a starting point, not proof: benchmark DARR against your own simple baselines before swapping anything into production. There is no Malaysia or Southeast Asia angle in the text.