IBM Granite TimeSeries TTM R2

TinyTimeMixers R2 is IBM's answer to a question the rest of the field answered with brute force: how large does a model have to be to forecast a number that changes over time — electricity demand next Tuesday, traffic on a junction, units sold in a warehouse. The industry's answer in 2024 was hundreds of millions to billions of parameters. IBM's is one million. The repository was pulled about 388,000 times in the thirty days to 1 September 2026, making it the third most downloaded model IBM publishes — ahead of the flagship 30-billion-parameter Granite chat model, and behind only two search encoders. It sits in production systems that never speak to a person. The design principle is the opposite of a general-purpose model. Rather than one large network covering every forecasting setting, IBM ships a family of small networks, each pre-trained for one specific pair of numbers: how much history it reads and how far ahead it predicts. The repository holds these variants as separate branches — 512, 1024 and 1536 points of history, forecasts of 96 to 720 points ahead — and a helper picks the matching one. The paper behind it was accepted at NeurIPS 2024. What makes the claim interesting is the comparison set. IBM reports that TTM beats TimesFM, Moirai, Chronos, Lag-Llama, Moment, GPT4TS, TimeLLM and LLMTime on zero-shot and few-shot forecasting benchmarks — a list that includes models built on language-model architectures with billions of parameters. If it holds, the whole premise that forecasting benefits from language-model scale is wrong, at least at these horizons. The numbers are the manufacturer's own; nobody has independently repeated the whole comparison. R2 was pre-trained on roughly 700 million samples, against 250 million for R1, and IBM puts the gain at over 15 percent on standard benchmarks. The later R2.1 branches raise that to about a billion samples and add daily and weekly resolutions. The honest limitation is in the card: the open version handles minute-to-hour data, and monthly or yearly series are out of scope, because the model needs at least 512 points of history. Padding short series with zeros to reach that length is explicitly discouraged. The licence is plain Apache 2.0. Fine-tuning to a specific dataset is expected rather than exceptional — IBM states that five percent of the usual training data is enough to be competitive, and that all of it runs on a single GPU or a laptop.

▸Videos

No videos yet.

▸Reviews