Primary use stated or inferred cautiously from official documentation.
IBM
IBM / ibm-granite/granite-timeseries-ensemble-r1
Granite Timeseries Ensemble R1
Time-series forecasting ensemble published by IBM on October 5, 2026, according to the official repository's public publication event. It combines PatchTST-FM-r1, PatchTST-FM-r2, FlowState-r1.1 and TTM-r3. It distributes reproducible recipes, not merged weights or a newly trained checkpoint: the four members are downloaded from their respective repositories at runtime.
01 / SUMMARY
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
02 / TECHNICAL SHEET
Limits and integration
| API ID | ibm-granite/granite-timeseries-ensemble-r1 |
|---|---|
| Model type | Time series |
| Access model | Open ensemble |
| License | Apache 2.0 · check member licenses |
| Deployment | Self-hosted with granite-tsfm ≥0.3.10 |
| Release | 05 OCT 2026 |
| Knowledge cutoff | Not published |
| Entrance | Numerical time series |
| Exit | Quantiles and numerical forecasts |
| Context window | Depends on each member model |
| maximum output | Nine quantiles from 0.1 to 0.9; median as the point forecast |
| Reasoning | Not applicable |
| Published tools | |
| Structured Outings | Tabular forecasts and quantiles |
| Batch processing | Not published |
| Prompt cache | Not published |
| Fine-tuning | Not published |
| Verified platforms | Hugging Face · Granite TSFM |
Zero-shot forecasting and uncertainty experiments for demand, energy or sensors. The default linear_pool recipe uniformly pools quantiles without learning weights; it returns nine quantiles from 0.1 to 0.9 and uses the median as the point estimate. The alternative iqr_weighted method favors narrow intervals and orders quantiles through isotonic regression.
Each multivariate channel is forecast independently: cross-channel relationships are not used. Loading four models increases memory and latency; context, horizon, missing-data and preprocessing limits depend on each member. A narrow interval does not guarantee accuracy or calibration. Validate on representative data and monitor distribution shifts before automating decisions.
03 / CAPABILITIES
What it can do
Orientation
Zero-shot forecasting and uncertainty experiments for demand, energy or sensors. The default linear_pool recipe uniformly pools quantiles without learning weights; it returns nine quantiles from 0.1 to 0.9 and uses the median as the point estimate. The alternative iqr_weighted method favors narrow intervals and orders quantiles through isotonic regression.
Context
Depends on each member model · Nine quantiles from 0.1 to 0.9; median as the point forecast
Tools and integration
Not published
Access
Hugging Face · separate recipes and checkpoints
Input modalities
Numerical time series
04 / PRICES
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | No published official API rate | Standard API rate |
| Cached input | Not published | Reading reused prefixes |
| Cache write or storage | Not published | The condition varies by provider |
| Standard output | Self-hosted infrastructure cost | May include reasoning tokens |
| Batch input | Not published | Asynchronous processing |
| Batch output | Not published | Asynchronous processing |
The identifier refers to a downloadable recipe, not a verified commercial API endpoint. There is no official per-token rate; keep cost unknown and budget for infrastructure. The recipe and implementation are Apache 2.0, but each checkpoint retains its own license. The expanded recipes include IBM Research members under separate, potentially non-commercial terms.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
05 / EVALUATIONS
How to read the results
“
IBM-published results, not reproduced by Inferama, over 97 GIFT-Eval configurations. MASE and CRPS/MWQL are normalized against Seasonal Naive and geometrically aggregated; lower is better. Ranks summarize positions per configuration, not percentages. Standings are dated September 28, 2026; the visual comparison filters licenses and other criteria, so it does not establish absolute leadership. Keep linear_pool and iqr_weighted separate.
| Benchmark | Result | Metric | Source |
|---|---|---|---|
| GIFT-Eval MASE | 0,6830 | Normalized MASE | View source ↗ |
| GIFT-Eval CRPS | 0,4707 | Normalized CRPS/MWQL | View source ↗ |
| GIFT-Eval MASE Rank | 37,19 | Aggregate rank | View source ↗ |
| GIFT-Eval CRPS Rank | 39,91 | Aggregate rank | View source ↗ |
| GIFT-Eval MASE | 0,6830 | Normalized MASE | View source ↗ |
| GIFT-Eval CRPS | 0,4700 | Normalized CRPS/MWQL | View source ↗ |
| GIFT-Eval MASE Rank | 37,65 | Aggregate rank | View source ↗ |
| GIFT-Eval CRPS Rank | 36,90 | Aggregate rank | View source ↗ |
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
06 / VERSIONS
Chronology
Ensemble R1 · linear_pool
Default recipe with four Granite members and uniform quantile pooling.
Ensemble R1 · iqr_weighted
Alternative granite-iqr recipe: same members; interquartile-range weighting.
07 / SOURCES