SSyncDev

ACE2.5-ERA5

SyncDev model directory · Profile updated Nov 21, 2026

Compare with another model

ACE2.5-ERA5 is Ai2's open stochastic climate emulator trained on ERA5, producing calibrated atmospheric ensembles from days to centuries at low compute.

Key specifications

Specification score
5.0
License
Apache-2.0

The specification score summarizes recorded features and limits. It is not a benchmark result.

Pricing conditions: Open weights; no hosted API.

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Text input

Produces

  • Text output

Capabilities

Open weights

What is ACE2.5-ERA5?

Climate experiments are usually limited by compute, not ideas. ACE2.5-ERA5, which Ai2 published on 19 November 2026, is a machine-learning emulator that steps the atmosphere forward from days to centuries and, unlike its predecessor ACE2-ERA5, does so stochastically: each run is one member of a calibrated ensemble rather than a single deterministic trajectory.

It is a research tool. Ai2 states explicitly that it should not be used for operational climate predictions.

Three checkpoints, three jobs

File Grid Step Parameters Best for
ace2.5-era5.tar 1° Daily 46M Default: climate skill with good throughput
ace2.5-era5-4deg.tar 4° Daily 14M Cheap long runs; 5× throughput, 10× lower training cost
ace2.5-era5-6h.tar 1° 6-hourly 46M Short-range weather and tropical cyclones

Each is seed 0. Extra seeds 1–3 exist for the two daily models, and a folder of ablation checkpoints reproduces the paper's figures (including an ACE2-ERA5 baseline), though those are not meant for downstream work. Training data are 1° and 4° regriddings of ERA5 reanalysis covering 1940–2025, published alongside the weights.

What changed from ACE2

Four improvements stand out:

  • Ensembles that mean something. Noise-conditioned layer normalisation, trained on CRPS plus a spectral energy score, gives spread that matches error to within 4% from day 3.
  • Rain that looks like rain. At wavenumbers 100–179 the model keeps 86% of ERA5's precipitation spectral power. ACE2-ERA5 kept 36%, which is why its fields looked smeared.
  • A tenth of the size. A spectral bottleneck cuts parameters from 456M to 46M.
  • Faster. The daily step and bottleneck together give 3.2× quicker inference.

One trade-off: the daily model's medium-range weather skill falls below ACE2-ERA5. The 6-hourly checkpoint closes that gap.

Where it is trustworthy, and where it isn't

Ai2 lists the strengths clearly. The model reproduces the atmospheric warming response to combined sea-surface temperature and CO₂ forcing over 1940–2020, the response to El Niño, the Madden–Julian Oscillation, a realistic stratospheric polar vortex, and (in the 6-hourly version) the geographic distribution of tropical cyclones. Global dry air mass and moisture are conserved exactly.

The weaknesses matter just as much. Separate sensitivities to SST and to CO₂ are not entirely realistic. Interannual variability of global-mean surface temperature is not calibrated. And it is not expected to generalise to large perturbations, so a doubled-CO₂ experiment is outside its envelope.

Trying it

Download a checkpoint, pip install fme, write an inference YAML naming the checkpoint, initial condition and forcing data, then run python -m fme.ace.inference <config.yaml>. Weights are Apache-2.0.

For a fair first test, hold out a recent decade, run a small ensemble from the 4° model and the 1° model, and compare precipitation spectra and ENSO response against reanalysis. If the coarse model is close enough for your question, you save an order of magnitude of compute. Ai2's Earth-observation work, such as OlmoEarth v1.3 Base, is listed in our model directory; general forecasters like TimesFM 3.0 sit under time-series forecasting but solve a different problem.

How to evaluate ACE2.5-ERA5

Compare the model on the work you need to complete.

Tasks and inputs

ACE2.5-ERA5 is listed for Data & control, Time Series Forecasting. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.

Choosing a deployment

The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

Start with ace2.5-era5.tar (1°, daily, 46M parameters). Use the 4° file for cheaper long runs and the 6-hourly file for short-range weather or tropical cyclones.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

Mistral AI

Timestral 1.0

Timestral 1.0 is Mistral's open time-series forecasting family in four sizes (5M-209M), producing zero-shot probabilistic forecasts on CPU.

32.77K context

allenai

HiRO-ACEv2

HiRO-ACEv2 is Ai2's open AI climate emulator that produces 3 km precipitation, winds and pressure from a 100 km run, over 300x cheaper than a numerical model.

— context

IBM

Granite TimeSeries PatchTST-FM-r2

Granite TimeSeries PatchTST-FM-r2 is IBM's open ~385M zero-shot forecasting model, dual-licensed Apache-2.0/OpenMDW, with 8,192-step context.

8.19K context

Google

TimesFM 3.0 PyTorch

TimesFM 3.0 is a numerical time-series forecasting checkpoint for PyTorch under a custom noncommercial licence.

— context