Best Time-Series Forecasting Models for Energy Load in 2026: timesfm-3.0-pytorch

Quick Answer

timesfm-3.0-pytorch (google) is the top pick for energy load forecasting as of September 2026 because it is the newest release; the ordering is newest release first, then downloads, with no public benchmark yet for any candidate [3] [5] [4] [10] [7] [9] [1]. In order, the ranking is timesfm-3.0-pytorch (google), granite-timeseries-patchtst-fm-r2 (ibm-granite), granite-timeseries-ttm-r3 (ibm-granite), chronos-2-small (autogluon), chronos-2 (amazon), chronos-2 (autogluon), and timesfm-2.5-200m-pytorch — established pick (google) [3][5].

Key Takeaways

• Google’s timesfm-3.0-pytorch (google) ranks first because its August 24, 2026 release is the newest eligible release; its energy-load benchmark status is no public benchmark yet. [3][5]
• Ranking note: No public benchmark scores any candidate, so the ordering is newest release first, then downloads; downloads only break ties. Scope admits only time-series forecasting models tagged time-series-forecasting from labs with a published paper or leaderboard record, or other publishers above 10,000 downloads in the last 30 days. [1][3][4][5][7][9][10]
• IBM Granite’s granite-timeseries-patchtst-fm-r2 (ibm-granite) has 385M parameters and 1.5 GB of F32 weights under openmdw-1.0. [5] IBM Granite’s granite-timeseries-ttm-r3 (ibm-granite) has 1M parameters under Apache-2.0; neither has a public benchmark yet. [4][5]
• AutoGluon’s chronos-2-small (autogluon) offers 28M parameters and 0.1 GB of F32 weights under Apache-2.0. [10] Amazon’s chronos-2 (amazon) and AutoGluon’s chronos-2 (autogluon) each have 119M parameters and 0.5 GB of F32 weights under Apache-2.0; all have no public benchmark yet. [7][9][10]
• Check licensing before choosing Google’s timesfm-3.0-pytorch (google): its license is timesfm-non-commercial-license-v1.0, and its F32 weights occupy 1.3 GB. [3] Treat weight size as a starting point for local hardware planning, not a verified runtime RAM or GPU requirement.
• Google’s google/timesfm-2.5-200m-pytorch is the established pick, with 231M parameters, 0.9 GB of F32 weights and Apache-2.0 licensing. Its September 2, 2025 release bypasses the 12-month gate because it is among its family’s three most-downloaded models; the exemption does not grant first place, and its benchmark status remains no public benchmark yet. [1]

How do local forecasting models compare on specifications, licenses and energy-load benchmarks?

Which local time-series forecasting models should you evaluate for energy load?

1. timesfm-3.0-pytorch

timesfm-3.0-pytorch from google ranks first under the ordering rule—newest release first, then downloads—with a publication date of August 24, 2026.[3][5] Its benchmark status is “no public benchmark yet,” so the position reflects release timing rather than demonstrated energy-load forecasting accuracy.

The model has 331M parameters and published F32 weights totaling 1.3 GB.[3] For local hardware planning, treat that weight size as a starting point, not a complete memory requirement. No specific GPU or system RAM requirement is established. Before choosing hardware, measure runtime memory and latency with your intended load histories and forecast horizons.

Use the model as a candidate for local energy-load evaluation, checking forecast errors on your own held-out demand data before deployment. The main caveat is its timesfm-non-commercial-license-v1.0 license; review those terms before considering a commercial application.[3]

2. granite-timeseries-patchtst-fm-r2

granite-timeseries-patchtst-fm-r2 from ibm-granite ranks second under the ordering “newest release first, then downloads”: its Hugging Face publication date is 2026-08-07, behind google’s timesfm-3.0-pytorch (google), published on 2026-08-24.[5][3] Its benchmark status is “no public benchmark yet,” so the position does not establish superior energy-load forecasting accuracy.

The model has 385M parameters, with F32 weights occupying 1.5 GB, and uses the openmdw-1.0 license.[5] For local hardware planning, treat that weight footprint as a starting point, not a complete memory requirement. A specific GPU or system-RAM capacity cannot be established from the weight file alone; measure total memory consumption with your intended workload before committing hardware.

Use it as a candidate for local energy-load evaluation when you want to assess the PatchTST approach, described in “A Time Series is Worth 64 Words: Long-term Forecasting with Transformers.”[6] The practical caveat is unverified task accuracy: validate forecasts on your own load history before selecting it for operational use.

3. granite-timeseries-ttm-r3

granite-timeseries-ttm-r3 by ibm-granite occupies the third position under the ranking’s “newest release first, then downloads” rule, based on its publication date of 2026-05-21.[4][5][10] Its position reflects release order; energy-load forecasting accuracy remains unverified, with no public benchmark yet.[4]

The model has 1M parameters and an Apache-2.0 license.[4] Weight-only storage at 4-bit precision would be 0.5 MB, estimated (params x 0.5 bytes) from the cited parameter count.[4] Treat that calculation as a storage estimate, not a verified runtime memory requirement or confirmation that quantized execution is supported. A specific CPU, GPU or system-RAM recommendation cannot be established from the available specifications.

Use the model as a compact candidate for local energy-load evaluation, with hardware sizing and forecast quality checked in your intended runtime. The practical caveat is the missing public benchmark: validate forecasts against your own load history before choosing it for operational use.

4. chronos-2-small

chronos-2-small from autogluon ranks fourth under the ordering of newest release first, then downloads, with a Hugging Face publication date of December 3, 2025.[10] Its position reflects release timing rather than demonstrated energy-load accuracy: no public benchmark yet establishes its performance against the other candidates.

The model has 28M parameters, published F32 weights of 0.1 GB, and an Apache-2.0 license.[10] For local deployment, treat the published weight size as a storage figure, not a total memory requirement. A validated CPU, GPU, or RAM minimum is not specified; measure runtime memory on your intended workload before selecting hardware.

Consider it for local energy-load forecasting trials when a compact, Apache-licensed model suits your deployment requirements.[10] Evaluate it on held-out load data at your intended forecast horizon, including peak-demand periods. The caveat is the missing public benchmark: its ranking does not establish forecasting quality for your workload.

5. chronos-2

chronos-2 by amazon ranks fifth under the ordering “newest release first, then downloads,” with a Hugging Face publication date of October 30, 2025.[7] Its energy-load status is “no public benchmark yet,” so the position does not establish forecasting accuracy. Evaluate it against your existing load forecast before adopting it.

The model has 119 million parameters, distributed as 0.5 GB of F32 weights, under the Apache-2.0 license.[7] For local hardware planning, use that weight footprint as a starting point. The weight size alone does not establish a total RAM requirement or justify a particular GPU recommendation; measure runtime memory with your intended workload.

A practical use is evaluating a locally hosted energy-load forecasting candidate with an Apache-2.0 license.[7] The caveat is the missing public benchmark: judge performance on your load history and forecast horizon before making an operational choice.

6. chronos-2

chronos-2 by autogluon (chronos-2 (autogluon)) ranks sixth under the ordering rule: newest release first, then downloads.[9] Its Hugging Face publication date is October 6, 2025, earlier than Amazon’s chronos-2 release on October 30, 2025.[7][9] The model has no public benchmark yet, so its position does not establish comparative energy-load forecasting accuracy.

The checkpoint contains 119 million parameters, with 0.5 GB of F32 weights, and uses the Apache-2.0 license.[9] Treat that weight size as a storage figure, not a complete memory budget. A specific CPU, GPU, RAM or VRAM requirement cannot be established from the published specifications; measure peak memory with your intended workload before choosing local hardware.

Use this checkpoint as a permissively licensed candidate for local energy-load evaluation.[9] Compare forecasts against your existing baseline on held-out load data, including peak-demand periods. The practical caveat is the absence of a public benchmark establishing its suitability for that task.

7. timesfm-2.5-200m-pytorch — established pick

timesfm-2.5-200m-pytorch — established pick by google ranks seventh under the ordering rule “newest release first, then downloads,” having first appeared on Hugging Face on September 2, 2025.[1] As one of its family’s three most-downloaded models, it bypasses the twelve-month recency gate as an established pick.[1] Its position does not establish an energy-load accuracy advantage: no public benchmark yet.

The model has 231M parameters, published F32 weights of 0.9 GB, and an Apache-2.0 license.[1] For local hardware planning, use that weight footprint as a starting point, then measure runtime memory with your intended workload. A specific GPU or total RAM requirement cannot be stated from the weight size alone.

A practical use is evaluating an Apache-2.0-licensed forecasting model against your existing energy-load baseline.[1] Check forecast quality and memory consumption on the machine intended for deployment. The caveat is benchmark coverage: treat this as an evaluation candidate, with energy-load suitability still to establish on your own data.

What hardware requirements can you estimate from forecasting model weights?

Forecasting model weights provide an estimated baseline for local storage and memory planning, but they do not establish a complete RAM requirement or a supported GPU configuration.

Google’s timesfm-3.0-pytorch (google) lists F32 weights of 1.3 GB.[3] IBM Granite’s granite-timeseries-patchtst-fm-r2 (ibm-granite) lists F32 weights of 1.5 GB.[5] Google’s google/timesfm-2.5-200m-pytorch lists F32 weights of 0.9 GB.[1] Treat those figures as weight footprints, rather than total inference memory requirements.

Amazon’s chronos-2 (amazon) and AutoGluon’s chronos-2 (autogluon) each list F32 weights of 0.5 GB.[7][9] AutoGluon’s chronos-2-small (autogluon) lists F32 weights of 0.1 GB.[10] IBM Granite’s granite-timeseries-ttm-r3 (ibm-granite) has 1M parameters; its reported F32 weight size is rounded to 0.0 GB, which should not be interpreted as zero memory usage.[4]

A practical hardware budget should leave room beyond the weights for the loaded application and forecast execution. No specific GPU model or total system RAM capacity can be justified from these weight figures alone. Avoid treating a weight footprint as a guarantee that a complete forecasting workload will fit.

For energy-load deployment, use the listed footprints to shortlist models for a local trial. Then measure memory use and inference time with your intended input length, forecast horizon and batch size before committing hardware. Weight size alone does not establish forecasting accuracy or serving speed.

Which forecasting model licenses allow commercial use?

The candidates licensed under Apache-2.0 or OpenMDW-1.0 allow commercial use; Google’s timesfm-3.0-pytorch (google) uses timesfm-non-commercial-license-v1.0 and does not provide that permission under its published license.[1][4][5][7][9][10][3]

The Apache-2.0 options are Google’s google/timesfm-2.5-200m-pytorch,[1] IBM Granite’s granite-timeseries-ttm-r3 (ibm-granite),[4] Amazon’s chronos-2 (amazon),[7] and AutoGluon’s chronos-2 (autogluon) and chronos-2-small (autogluon).[9][10] For an energy-load forecasting service, an internal business application, or a customer deployment, those models provide a commercially usable licensing starting point. Commercial permission still means following the applicable license terms; treat the license review as part of deployment.

IBM Granite’s granite-timeseries-patchtst-fm-r2 (ibm-granite) uses OpenMDW-1.0, providing another commercially usable option.[5] Review its own terms when preparing a deployment or distributing model weights. Avoid assuming that the compliance steps for an Apache-licensed model automatically cover a model distributed under a different license.

Google’s newer TimesFM release needs particular attention: timesfm-3.0-pytorch (google) carries a non-commercial license,[3] whereas google/timesfm-2.5-200m-pytorch carries Apache-2.0.[1] A model upgrade within the same family can therefore change the licensing basis of a commercial deployment. Record the exact repository and license alongside the weights you deploy, and make license compatibility a selection requirement before evaluating forecasting accuracy or hardware needs.

How should you evaluate energy-load forecasting models with no public benchmark yet?

Evaluate energy-load forecasting models with no public benchmark yet through backtests on your own load data, then measure local runtime, memory use and deployment suitability.

Define the forecast horizon, update schedule and inputs before comparing candidates. Use chronological holdouts and repeat evaluation across successive forecast dates. Include seasonal changes, holidays and periods of unusual demand. Keep preprocessing and model selection confined to training data, and supply weather forecasts available at prediction time rather than subsequently observed weather.

Compare each candidate against a seasonal baseline that repeats the corresponding historical load pattern. Report absolute error alongside errors during peak demand and systematic overprediction or underprediction. Break results out by forecast horizon and operating conditions. Choose acceptance criteria around the decisions the forecast supports, such as scheduling capacity or managing demand.

Run each evaluation on the intended local machine. Record inference time, peak memory, preprocessing overhead and failed forecasts using consistent workloads. Google’s timesfm-3.0-pytorch (google) lists F32 weights of 1.3 GB;[3] treat that as weight storage, and measure total runtime memory separately before choosing hardware.

Check licensing before deployment: Google’s timesfm-3.0-pytorch (google) uses timesfm-non-commercial-license-v1.0.[3] Document the checkpoint, evaluation dates, data windows and configuration so colleagues can reproduce the comparison. Treat release recency as a shortlist criterion; require measured performance on your load data before selecting a model.

Frequently Asked Questions

Which model ranks first for local energy-load forecasting?

Google’s TimesFM 3.0 (timesfm-3.0-pytorch (google)) ranks first because its Hugging Face publication date is 2026-08-24, later than the next candidate’s 2026-08-07 publication date.[3][5] The position reflects release recency, not demonstrated energy-load accuracy: no public benchmark yet. Its timesfm-non-commercial-license-v1.0 license also matters when choosing a model for an operational forecasting project.[3]

How is the ranking ordered?

The ordering is newest release first, then downloads; no public benchmark scores any candidate. Scope: this ranking admits only time series forecasting models from labs with a published paper or leaderboard record, or other publishers above 10,000 downloads in the last 30 days, using Hugging Face’s time-series-forecasting pipeline tag.[1][3][4][5][7][9][10] Google’s TimesFM 2.5 (google/timesfm-2.5-200m-pytorch) is an established pick: its family popularity permits a recency exemption, but ordering follows the same rule and the exemption cannot secure first place.[1]

Which licenses should I check before using a model commercially?

Google’s TimesFM 3.0 uses timesfm-non-commercial-license-v1.0.[3] IBM Granite’s granite-timeseries-patchtst-fm-r2 uses openmdw-1.0.[5] IBM Granite’s granite-timeseries-ttm-r3 and Google’s TimesFM 2.5 use Apache-2.0.[4][1] Start your deployment review with the exact license attached to the selected repository. An open-weight release should not substitute for checking whether the license permits your intended energy-load forecasting service.

How much memory do these models need to run locally?

Published weight sizes provide a starting point, not a complete runtime memory requirement. TimesFM 3.0 lists F32 weights of 1.3 GB, while granite-timeseries-patchtst-fm-r2 lists 1.5 GB.[3][5] The compact granite-timeseries-ttm-r3 lists 1M parameters.[4] A specific GPU or RAM recommendation would require an estimate or a measured configuration. Measure memory with your intended context, forecast horizon and batch size before selecting hardware.

Which Chronos release should I evaluate first?

Under the release-first ordering, start with AutoGluon’s chronos-2-small (chronos-2-small (autogluon)), published 2025-12-03 with 28M parameters and 0.1 GB of F32 weights.[10] Amazon’s Chronos-2 (chronos-2 (amazon)) follows, published 2025-10-30; AutoGluon’s chronos-2 (chronos-2 (autogluon)) was published 2025-10-06.[7][9] Both larger entries list 119M parameters and Apache-2.0, as does the small release for its license.[7][9][10] The ordering does not establish an energy-load accuracy advantage.

Are there dated energy-load benchmarks that establish an accuracy winner?

No public benchmark yet scores these candidates against one another, so the ranking cannot establish an energy-load accuracy winner. The model-card snapshot is dated 2026-09-25; that is a documentation date, not an energy-load benchmark date.[1][3][4][5][7][9][10] For deployment, compare candidates on your own held-out load history using the same forecast horizons and available inputs. Record forecast error, peak-load misses, runtime and memory alongside the evaluation dates.

Sources

1. google/timesfm-2.5-200m-pytorch model card (Hugging Face) — 2026-09-25
2. A decoder-only foundation model for time-series forecasting — 2023-10-14
3. google/timesfm-3.0-pytorch model card (Hugging Face) — 2026-09-25
4. ibm-granite/granite-timeseries-ttm-r3 model card (Hugging Face) — 2026-09-25
5. ibm-granite/granite-timeseries-patchtst-fm-r2 model card (Hugging Face) — 2026-09-25
6. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers — 2022-11-27
7. amazon/chronos-2 model card (Hugging Face) — 2026-09-25
8. Chronos: Learning the Language of Time Series — 2024-03-12
9. autogluon/chronos-2 model card (Hugging Face) — 2026-09-25
10. autogluon/chronos-2-small model card (Hugging Face) — 2026-09-25
11. Chronos-2: From Univariate to Universal Forecasting — 2025-10-17