AUREON

EARTH OBSERVATION / OPEN-SOURCE INFERENCE

Inference for
specialized intelligence.

Open-source inference for
multimodal and generative Earth models.

View on GitHub

From model to application.

The AUREON engine

Built around the model.

Inputs, preprocessing and precision defined for each model family.

View model families

One clear interface.

Submit jobs. Track progress. Retrieve georeferenced outputs.

View API reference

The locked local v0.1.0 package integrates scene, HTTP and Cloud features directly with the resident Best runtime. One installed Engine; no model subprocess adapter. Original benchmark scope and GitHub downloads remain unchanged; hosted inference and PyPI remain closed.

Built for the inference path.

How it works

01 / RESIDENT

Load once.

Keep model weights resident across jobs, with request-owned sampling state.

02 / REUSE

Reuse eligible work.

Reuse unchanged conditioning only when identity and recipe checks allow it.

03 / EXECUTE

Replay eligible graphs.

Use guarded CUDA graphs in TerraMind decoding, with native execution available.

04 / DELIVER

Keep the geography.

Export verified COGs with grids, masks and provenance.

Implemented for the recorded TerraMind path. Each adapter owns its model-specific execution.

More inference from your GPU.

TerraMind 1.0 base · 224 × 224 S2L2A → synthetic S1GRD · FP32

MetricTerraTorch + native vLLMAUREONImprovement
Queued throughput1.86670 jobs/s2.34980 jobs/s25.88% higher
End-to-end latency, sequential median617.43 ms446.40 ms27.70% lower
End-to-end latency, burst median1,102.20 ms852.64 ms22.64% lower
Warm worker-hours / 1,000 jobs0.148807 h0.118213 h20.56% lower

Tested on RTX 3070 · 8 GB VRAM

Historical Best · 27 Oct 2026 · measured wheel 0fa5ef2b. These figures are not a benchmark of the current integrated e09e710c package. Warm queued work, excluding startup; compare cost at the same worker-hour rate. Methodology and scope

AUREON USE CASE / LAND-COVER SEGMENTATION

Land-cover mapping with AUREON.

Recorded result

Sentinel-2 imagery to a class-map COG. This recorded workflow uses KAO-TerraMind B.

POST /v1/jobs
{
  "scene_id": "your-registered-scene"
}
COMPLETED JOB
{
  "status": "completed",
  "output": "class_map.tif"
}

Tested contract illustration · Hosted jobs not yet open

View the workflow
Audited Sentinel-2 RGB scene from central Thailand
Sentinel-2 inputRecorded Thailand scene
Recorded class prediction for the same Sentinel-2 scene
Class mapRecorded output · Same source grid

MODEL FAMILIES

Explicit routes.
Room to grow.

Supported means a defined, tested workflow. New adapters and modality routes are qualified separately.

Supported · Published alpha

KAO-TerraMind

Input
Sentinel-2 L2A · 12 bands
Validated task
10-class segmentation
Example output
Class-map COG + validity mask

B: local scene inference and recorded COGs. A/C retain their separately qualified full-reference scope.

View family details
Supported · Published alpha

TerraMind

Input
Sentinel-2 L2A · native tiles and prepared scenes
Validated task
S2L2A → synthetic S1GRD
Output
Continuous VV/VH COG + provenance

TerraMind 1.0 base. Synthetic output is not a sensor measurement.

View family details

More interfaces. Explicit qualification.

Development · v0.1.0
Available interfaces · CPU checks

From STAC to outputs.

Selected GeoTIFF windows connect to the existing TerraMind scene engine and COG delivery. Installed CPU fixtures pass; no new model or performance qualification.

Workflow scope
Experimental · Real-model acceptance

COP-GEN.

One tested route: a 192 × 192 Sentinel-2 crop to synthetic S1RTC. The installed Torch 2.10 SDK matches its native reference. Interface included; prior-artifact acceptance remains separate.

Scope and limitations

GitHub v0.1.0 now downloads the accepted Best package e09e710c. Historical benchmark figures remain tied to their original measured archive. Hosted inference and PyPI remain closed.

Build for
Earth observation.

Explore the docs and recorded results.

Contact usRead the docs