A model is still running

A model is still running or downloading. Leaving will stop the current operation. Stay on this page to let it finish.

Skip to content
LocalAI.toolsBeta

Join the waitlist

Get an email when the new apps and model runs launch.

No account needed.

OlmoEarth 1.3 Base

Encode one32×32 Sentinel2 L2A multispectral satellite patch into complete spatial embeddings and an optional relative magnitude map.

Browser inferencePowered by WebAI · Private · Unlimited runs
About this model

Encode one32×32 Sentinel2 L2A multispectral satellite patch into complete spatial embeddings and an optional relative magnitude map.

Model released
Unknown
Source repository created
Sep 4, 2026
Model weights · FP32
412.3 MB

Best for

  • Satellite patch similarity and retrieval experiments

Keep in mind

  • Only Sentinel2 L2A has measured task quality. Sentinel1 and Landsat remain unqualified.
  • Two held-out patches from one acquisition establish narrow retrieval evidence only; no land-cover classifier or broad accuracy claim.
  • Requires exact sensor bands, raw satellite values and calendar dates; RGB photographs are not valid input.
  • Artifact license restricts military, surveillance and extractive uses; retain the full terms.
  • Current Mac public API/lifecycle/sample-array parity passes. Production app, complete backup and cold source recovery remain pending.
  • The atomic encoder returns its tensors together; the pinned runtime has no incremental task output. Streaming is unavailable.

Processed on your deviceInference runs in this browser. Input/output sharing requires consent for each run. Basic run logs include device information and approximate connection location. Model and runtime files download from external hosts. Saved downloads may be cleared by your browser.

Sources & technical details
Original model
https://huggingface.co/allenai/OlmoEarth-v1_3-Base
Release verification
Unknown
Release evidence
Exact checkpoint release day not established by bounded official review; repository creation is distinct.
Source updated
Oct 13, 2026
Browser model files
https://assets.localai.tools/models/converted/v1/olmoearth-v1-3/30ea64d178d08192578a5ba6dd0eeda93b29bec1
Files updated
Nov 16, 2026
Artifact revision
30ea64d178d08192578a5ba6dd0eeda93b29bec1

Input

TEXT

Settings

Relative embedding magnitude, not land-cover labels. Adds a relative magnitude map alongside all native arrays. This does not assign land-cover labels.
Advanced settings 1
Spatial patch edge. Request token policy remains 1024. Selects the spatial embedding grid; small patches produce more tokens.

Output

Your result will appear here

Add input or try the sample, then run the model.

Benchmark evidence

Saved worker evidence is separate from benchmarks run on your device. It does not guarantee speed or quality on every browser.

OlmoEarth 1.3 Base

4 configurations tested
Passed4
Failed0
TestedNov 16, 2026
Runtimeonnxruntime-web 1.31.0-dev.20260914-8d85527a0

Quality criteria

{ "frozenBeforeEmbeddings": true, "task": "same-acquisition patch embedding retrieval", "reference": "ESA Sentinel-2 L2A SCL majority class, nearest-resampled independently from the input bands", "classIds": { "water": 6, "vegetation": 4 }, "minimumLabelPurity": 0.7, "training": [ "water-train", "vegetation-train" ], "heldOut": [ "water-test", "vegetation-test" ], "metric": "cosine nearest prototype on spatial-mean student_location_tokens", "requiredAccuracy": 1, "requiredCpuMpsCosine": 0.999, "limitations": "Two test patches from one Seattle acquisition, not independent regional/year/domain validation or a land-cover classifier. Training/test windows are spatially non-overlapping. Broader retrieval quality and other sensor modes remain unmeasured." }

PrecisionDeviceResult
FP32CPU · WASMPassedVery fast
FP32CPU · WASMPassedVery fast
FP32CPU · WASMPassedVery fast
FP32CPU · WASMPassedVery fast
Test setup & sample details
Quality gate
{ "frozenBeforeEmbeddings": true, "task": "same-acquisition patch embedding retrieval", "reference": "ESA Sentinel-2 L2A SCL majority class, nearest-resampled independently from the input bands", "classIds": { "water": 6, "vegetation": 4 }, "minimumLabelPurity": 0.7, "training": [ "water-train", "vegetation-train" ], "heldOut": [ "water-test", "vegetation-test" ], "metric": "cosine nearest prototype on spatial-mean student_location_tokens", "requiredAccuracy": 1, "requiredCpuMpsCosine": 0.999, "limitations": "Two test patches from one Seattle acquisition, not independent regional/year/domain validation or a land-cover classifier. Training/test windows are spatially non-overlapping. Broader retrieval quality and other sensor modes remain unmeasured." }
Tested
2026-11-16T17:44:31.446Z
Environment
Apple M3 Max/macOS27.2; exact pinned runtime; existing mirror active. Public API timings are scoped observations; retained standalone quality/speed evidence keeps original timestamps.
Runtime
onnxruntime-web 1.31.0-dev.20260914-8d85527a0
Artifact revision
30ea64d178d08192578a5ba6dd0eeda93b29bec1
Fixture
{
  "file": "model-samples/olmoearth-v1-3/water-test-browser-input.json",
  "size": 53465,
  "sha1": "8b6b9b3103bb6d00dd74231fb365e6f234066732",
  "sha256": "513eea1bcc3af5aedb990b780b140f13cfc8f44b421fb288b6aa489d4c2eb336",
  "url": "https://assets.localai.tools/samples/v1/files/513eea1bcc3af5aedb990b780b140f13cfc8f44b421fb288b6aa489d4c2eb336/water-test-browser-input.json"
}
Rerun when
artifact revision; runtime version; normalization; fixture checksum; quality gate; backend behavior
Per-case quality & failure details
[
  {
    "precision": "fp32",
    "device": "wasm",
    "status": "pass",
    "quality": {
      "passed": true,
      "metric": "finite prototype embedding",
      "score": 1,
      "scope": "Reference prototype construction; not a held-out prediction."
    },
    "runtime": {
      "durationMs": 42.625,
      "loadMs": 4678.80000000447,
      "inferenceMs": 42.75500000268221,
      "version": "1.31.0-dev.20260914-8d85527a0",
      "engine": "onnxruntime-web",
      "artifactRevision": "30ea64d178d08192578a5ba6dd0eeda93b29bec1",
      "grade": "good",
      "totalMs": 4721.555000007153,
      "pixelsPerSecond": 23950.415154619575,
      "actualDevice": "wasm",
      "actualPrecision": "fp32",
      "neuralExecuted": true
    }
  },
  {
    "precision": "fp32",
    "device": "wasm",
    "status": "pass",
    "quality": {
      "passed": true,
      "metric": "finite prototype embedding",
      "score": 1,
      "scope": "Reference prototype construction; not a held-out prediction."
    },
    "runtime": {
      "durationMs": 22.42000000178814,
      "loadMs": 4678.80000000447,
      "inferenceMs": 22.44999999552965,
      "version": "1.31.0-dev.20260914-8d85527a0",
      "engine": "onnxruntime-web",
      "artifactRevision": "30ea64d178d08192578a5ba6dd0eeda93b29bec1",
      "grade": "good",
      "totalMs": 4701.25,
      "pixelsPerSecond": 45612.472169438915,
      "actualDevice": "wasm",
      "actualPrecision": "fp32",
      "neuralExecuted": true
    }
  },
  {
    "precision": "fp32",
    "device": "wasm",
    "status": "pass",
    "quality": {
      "passed": true,
      "metric": "held-out same-acquisition retrieval accuracy",
      "score": 1,
      "scope": "Narrow two-class same-acquisition suite; exact native output parity separately passed. Training prototypes are reference construction, not independent held-out cases."
    },
    "runtime": {
      "durationMs": 19.814999997615814,
      "loadMs": 4678.80000000447,
      "inferenceMs": 19.83500000089407,
      "version": "1.31.0-dev.20260914-8d85527a0",
      "engine": "onnxruntime-web",
      "artifactRevision": "30ea64d178d08192578a5ba6dd0eeda93b29bec1",
      "grade": "good",
      "totalMs": 4698.635000005364,
      "pixelsPerSecond": 51625.913786430196,
      "actualDevice": "wasm",
      "actualPrecision": "fp32",
      "neuralExecuted": true
    }
  },
  {
    "precision": "fp32",
    "device": "wasm",
    "status": "pass",
    "quality": {
      "passed": true,
      "metric": "held-out same-acquisition retrieval accuracy",
      "score": 1,
      "scope": "Narrow two-class same-acquisition suite; exact native output parity separately passed. Training prototypes are reference construction, not independent held-out cases."
    },
    "runtime": {
      "durationMs": 22.28999999910593,
      "loadMs": 4678.80000000447,
      "inferenceMs": 22.309999994933605,
      "version": "1.31.0-dev.20260914-8d85527a0",
      "engine": "onnxruntime-web",
      "artifactRevision": "30ea64d178d08192578a5ba6dd0eeda93b29bec1",
      "grade": "good",
      "totalMs": 4701.109999999404,
      "pixelsPerSecond": 45898.70014489202,
      "actualDevice": "wasm",
      "actualPrecision": "fp32",
      "neuralExecuted": true
    }
  }
]

Something not quite right?

Tell us what happened with OlmoEarth 1.3. Loading, results, or the page itself—we’re here for all of it.

Attachments Optional

Up to 3 files · 50 MB each

Browser, OS, CPU/GPU, memory, screen, storage, and network details are included automatically. Model inputs, results, and browsing history are not collected.