IBM's 385M forecasting model is first among permissive licences

Granite Time Series PatchTST-FM-r2 builds its backbone from conformer blocks, which combine multi-head self-attention with a temporal convolution. IBM's stated reason for the pairing is to capture long-range and short-range temporal structure at once.

Attention gives every patch a path to every other patch, and the work grows with the square of the sequence. A convolution carries a fixed local receptive field and no such cost. Put both in a block and the convolution covers the neighbourhood, so attention depth is not spent rediscovering that adjacent patches correlate, and attention is left to work on patches that sit far apart. IBM calls the two mechanisms complementary.

What the release does not carry is an ablation. r2 is ranked on GIFT-Eval alongside TimesFM-3 and alongside the pretrained baselines, so the architectures are compared against each other on one evaluation. What is missing is the table that separates what the conformer block contributes from what the larger pretraining corpus contributes — same data, same compute, backbone swapped. That is the experiment I would want and have not seen. The weights, the architecture, the inference pipeline and the code that reproduces the benchmark numbers are all published, so it is one someone could run this week.

The post lists the larger pretraining corpus, the 99-quantile prediction head and imputation support as things r2 brings. It does not say which of the three r1 already had. Roughly 385M parameters. Context length up to 8,192. Dual-licensed under Apache-2.0 and OpenMDW-1.0, and you pick either.

Two licences and your choice reads as generous, and the post does not explain why you would reach for one over the other, or what OpenMDW-1.0 asks of you in exchange. That is a question for your own legal review rather than for the model card.

The ranking claim needs parsing, because "SOTA" in the announcement carries more weight than the leaderboard does.

GIFT-Eval splits its leaderboard into categories, and IBM is claiming one: zero-shot, replicable, no test leakage. PatchTST-FM-r2 is first in that category among permissively licensed models and second overall, behind TimesFM-3. IBM does not say what licence TimesFM-3 ships under. That gap is what the phrasing is built on. "First among models you can ship" is a narrower claim than "first", and for commercial work it is usually the only one of the two you can act on.

You are adding forecasting to a product and you want to avoid a legal review. Filter the leaderboard to zero-shot, replicable and permissively licensed, and there is one model to look at. Drop the licence filter and you are choosing between r2 and whichever terms TimesFM-3 ships under, about which IBM says nothing. Widen it again to the pool where pretrained models are added to the comparison, and by IBM's account r2 stays near the top of what remains. Three readings, all correct, off one leaderboard. The category filter matters more than the rank.

IBM reports geometric-mean CRPS of 0.467 and MASE of 0.6846 for the first of those positions. An aggregate over someone else's task mix is weak evidence about your series. What it does have going for it is that GIFT-Eval's replicable category requires the evaluation to be reproducible without test leakage, and IBM published the pipeline and code alongside the weights. The number is one you can re-run, which is more than most blog benchmarks offer. Re-running it still will not tell you how the model behaves on your data, only that the claim is checkable.

Context length up to 8,192. A 99-quantile prediction head, so the forecast is a distribution rather than a point. Imputation of missing values. If your history is longer than you can currently push through in one pass, the context buys you fewer chunks. If you feed a capacity or inventory system, a quantile head is the shape that system already wants, and you are not bolting a second model onto the output to produce intervals. If your telemetry arrives with holes, you do not have to backfill before the series goes in. Each of those depends on your sampling rate and your pipeline rather than on the model card. IBM also points at Confluent for running the Granite time-series family in streaming production settings, which is a pointer rather than a benchmark.

What I could not find in the release: latency, memory or cost per forecast; the hardware any of it was measured on; and results at horizons or sampling rates that look nothing like demand, prices, energy load, traffic or telemetry. Those are the numbers that decide a deployment. A leaderboard entry is an average over somebody else's task mix at somebody else's horizon, and the average can be strong while your slice of it is not. Without a per-task breakdown, the only way to find out whether the conformer block is doing anything for your data is to fine-tune r2 and a plain transformer on the same history and compare. The code to do that is published. The compute is yours.

The ranks are dated September 8, and they move.