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.

LightOnOCR 3

Transcribe still document images and PDF pages to Markdown or extract labeled regions, chart tables and grounded text.

Desktop inferencePrivate · Runs on your computer
About this model

Transcribe still document images and PDF pages to Markdown or extract labeled regions, chart tables and grounded text.

Created by
LightOn AI
License
Apache-2.0
Model released
Oct 8, 2026
Source repository created
Oct 7, 2026
Model weights · FP32
2.0 GB

Best for

  • Local document OCR with manual verification
  • Reading bounded scientific excerpts and simple scanned pages
  • Reproducible exact-checkpoint OCR experiments

Keep in mind

  • Accuracy scope covers saved English scientific, receipt and PDF fixtures. Review chart-derived numbers and OCR output against the original.
  • PNG/JPEG or PDF bytes up to 32 MiB. PDF rendering supports up to eight pages. Audio and video are unavailable.
  • Grounding boxes use normalized 0–1000 coordinates; output can be truncated by the token budget.

Processed on your deviceYour input is processed locally in the desktop app. Model and runtime files download from external hosts.

Target platforms · verification required

  • macos · aarch64, x86_64
  • windows · x86_64

Recommended configuration

FP32 · mps (verified platform only)

Sources & technical details
Original model
https://huggingface.co/lightonai/LightOnOCR-3-1B
Release verification
Verified
Release evidence
The official October 8 announcement lists all three exact checkpoints.
Source updated
Oct 9, 2026
Native model files
lightonai/LightOnOCR-3-1B
Files updated
Oct 9, 2026
Artifact revision
23d544085c886bbdace647bcfab91203299417e2

Desktop app

Sample result

LightOnOCR 3 1B

Run LightOnOCR 3 on your computer

Download the desktop app to load this model, provide input and view results locally.

Public desktop downloads are coming soon.

Device requirements

macOS Apple SiliconCPU / FP32Tested16.0 GiBRAM estimate
macOS Apple SiliconMPS / FP32Tested24.0 GiBRAM estimate
If your device falls short

MemoryAdvisory only: close other apps, reduce image/tile/token/batch settings or choose a smaller verified model/precision. CPU can need more RAM than GPU.

GPUIf the selected GPU runtime fails, existing CPU fallback may recover when enabled, but CPU has its own memory/runtime needs.

PlatformUse a compatible OS/architecture/runtime or another model. These estimates do not add hardware detection or block runs.

Benchmark evidence

Saved desktop measurements apply to the recorded hardware and runtime. Fallback cases report the device that actually ran; they do not prove GPU support on the requested device.

LightOnOCR 3 1B

2 configurations tested
Passed2
Failed0
TestedNov 14, 2026
Runtimetransformers 5.22.1

Quality criteria

Saved paragraph recall, chart location and completed table, receipt text, ordered PDF, EOS and per-request speed limit.

PrecisionDeviceResultPlatformInference / load
FP32mpsAcceleratedPassedNot measuredmacos · aarch643325 ms / 7209 ms
FP32cpuPassedNot measuredmacos · aarch648685 ms / 9501 ms
Test setup & sample details
Quality gate
Saved paragraph recall, chart location and completed table, receipt text, ordered PDF, EOS and per-request speed limit.
Tested
2026-11-14T11:41:28.886261+00:00
Environment
Apple M3 Max macOS 27.2; Transformers 5.22.1 and PyTorch 2.10.0
Runtime
transformers 5.22.1
Artifact revision
23d544085c886bbdace647bcfab91203299417e2
Fixture
{
  "file": "model-samples/desktop/lightonocr2/fixtures/scientific-paragraph.png",
  "url": "https://assets.localai.tools/samples/v1/files/de0959e7c3fe59f4d706bb84d993cd1aae000a6b25bcef96f472d3a675871dfb/scientific-paragraph.png",
  "size": 26674,
  "sha1": "b4e9dd3ea6464bc8e1ff1f95134eccfc25622d52",
  "sha256": "de0959e7c3fe59f4d706bb84d993cd1aae000a6b25bcef96f472d3a675871dfb"
}
Rerun when
artifact revision/weight change; pinned runtime/prediction/preprocessing change; fixture/checksum or quality-gate change; relevant host/backend behavior change
Per-case quality & failure details
[
  {
    "precision": "fp32",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "content": true,
      "eos": true,
      "speed": true,
      "metric": "reference-token-recall",
      "score": 0.9,
      "minimum": 0.9
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "Vision/language parameters, inputs and vision activations on selected device; rendering, tokenizer and decoding on CPU",
      "activationPlacement": {
        "languageOutput": "cpu",
        "visionOutput": "cpu"
      },
      "allPagesReachedEos": true,
      "artifactRevision": "23d544085c886bbdace647bcfab91203299417e2",
      "cpuThreads": 4,
      "decoding": "greedy",
      "device": "cpu",
      "durationMs": 8796.327915999998,
      "engine": "transformers",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "lightonocr-3-native",
        "defaultModelId": "lightonocr-3-0.8b-native",
        "displayName": "LightOnOCR 3",
        "id": "lightonocr-3",
        "variantId": "1b",
        "variantName": "1B"
      },
      "inferenceMs": 8684.54175000079,
      "inputKind": "image",
      "loadMs": 9500.678125,
      "maxNewTokens": 1024,
      "memoryMeasurementScope": "Process lifetime peakRSS through complete output construction; not a validated minimum.",
      "modelId": "lightonocr-3-1b-native",
      "neuralInferencePerformed": true,
      "outputLimitReached": false,
      "outputTokens": 73,
      "pageInferenceMs": [
        8684.54175000079
      ],
      "pagesPerSecond": 0.11514712333554146,
      "parameterDtype": "torch.float32",
      "pdfScale": 1.5,
      "peakResidentMemoryBytes": 6492241920,
      "placement": {
        "cpu/torch.float32": 532
      },
      "precision": "fp32",
      "provider": "PyTorch CPU",
      "requestedDevice": "cpu",
      "requestedPrecision": "fp32",
      "reused": false,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "tokensPerSecond": 8.405740003494527,
      "torchVersion": "2.10.0",
      "transformersVersion": "5.22.1",
      "version": "5.22.1",
      "visualizationMs": 2.427041996270418,
      "totalMs": 18185.21987500079,
      "grade": "interactive"
    }
  },
  {
    "precision": "fp32",
    "device": "mps",
    "status": "pass",
    "quality": {
      "passed": true,
      "content": true,
      "eos": true,
      "speed": true,
      "metric": "reference-token-recall",
      "score": 0.9,
      "minimum": 0.9
    },
    "runtime": {
      "acceleration": "partial",
      "accelerationDetails": "Vision/language parameters, inputs and vision activations on selected device; rendering, tokenizer and decoding on CPU",
      "activationPlacement": {
        "languageOutput": "mps:0",
        "visionOutput": "mps:0"
      },
      "allPagesReachedEos": true,
      "artifactRevision": "23d544085c886bbdace647bcfab91203299417e2",
      "cpuThreads": 4,
      "decoding": "greedy",
      "device": "mps",
      "driverAllocatedMemoryBytes": 5118115840,
      "durationMs": 3409.048792,
      "engine": "transformers",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "lightonocr-3-native",
        "defaultModelId": "lightonocr-3-0.8b-native",
        "displayName": "LightOnOCR 3",
        "id": "lightonocr-3",
        "variantId": "1b",
        "variantName": "1B"
      },
      "inferenceMs": 3325.1684170099907,
      "inputKind": "image",
      "loadMs": 7208.726209,
      "maxNewTokens": 1024,
      "memoryMeasurementScope": "Process lifetime peakRSS through complete output construction; not a validated minimum.",
      "modelId": "lightonocr-3-1b-native",
      "neuralInferencePerformed": true,
      "outputLimitReached": false,
      "outputTokens": 73,
      "pageInferenceMs": [
        3325.1684170099907
      ],
      "pagesPerSecond": 0.30073664686710977,
      "parameterDtype": "torch.float32",
      "pdfScale": 1.5,
      "peakResidentMemoryBytes": 6492389376,
      "placement": {
        "mps:0/torch.float32": 532
      },
      "precision": "fp32",
      "provider": "Metal/MPS",
      "requestedDevice": "mps",
      "requestedPrecision": "fp32",
      "reused": false,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "tokensPerSecond": 21.953775221299015,
      "torchVersion": "2.10.0",
      "transformersVersion": "5.22.1",
      "version": "5.22.1",
      "visualizationMs": 3.9316660258919,
      "totalMs": 10533.894626009991,
      "grade": "interactive"
    }
  }
]

LightOnOCR 3 0.8B

1 configurations tested
Passed1
Failed0
TestedNov 14, 2026
Runtimetransformers 5.22.1

Quality criteria

Saved paragraph recall, chart location and completed table, receipt text, ordered PDF, EOS and per-request speed limit.

PrecisionDeviceResultPlatformInference / load
FP32mpsAcceleratedPassedNot measuredmacos · aarch643809 ms / 4822 ms
Test setup & sample details
Quality gate
Saved paragraph recall, chart location and completed table, receipt text, ordered PDF, EOS and per-request speed limit.
Tested
2026-11-14T11:39:56.155837+00:00
Environment
Apple M3 Max macOS 27.2; Transformers 5.22.1 and PyTorch 2.10.0
Runtime
transformers 5.22.1
Artifact revision
36d8636925327c671bd8e479c517a004f7566bd9
Fixture
{
  "file": "model-samples/desktop/lightonocr2/fixtures/scientific-paragraph.png",
  "url": "https://assets.localai.tools/samples/v1/files/de0959e7c3fe59f4d706bb84d993cd1aae000a6b25bcef96f472d3a675871dfb/scientific-paragraph.png",
  "size": 26674,
  "sha1": "b4e9dd3ea6464bc8e1ff1f95134eccfc25622d52",
  "sha256": "de0959e7c3fe59f4d706bb84d993cd1aae000a6b25bcef96f472d3a675871dfb"
}
Rerun when
artifact revision/weight change; pinned runtime/prediction/preprocessing change; fixture/checksum or quality-gate change; relevant host/backend behavior change
Per-case quality & failure details
[
  {
    "precision": "fp32",
    "device": "mps",
    "status": "pass",
    "quality": {
      "passed": true,
      "content": true,
      "eos": true,
      "speed": true,
      "metric": "reference-token-recall",
      "score": 0.95,
      "minimum": 0.9
    },
    "runtime": {
      "acceleration": "partial",
      "accelerationDetails": "Vision/language parameters, inputs and vision activations on selected device; rendering, tokenizer and decoding on CPU",
      "activationPlacement": {
        "languageOutput": "mps:0",
        "visionOutput": "mps:0"
      },
      "allPagesReachedEos": true,
      "artifactRevision": "36d8636925327c671bd8e479c517a004f7566bd9",
      "cpuThreads": 4,
      "decoding": "greedy",
      "device": "mps",
      "driverAllocatedMemoryBytes": 3708944384,
      "durationMs": 3869.113917,
      "engine": "transformers",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "lightonocr-3-native",
        "defaultModelId": "lightonocr-3-0.8b-native",
        "displayName": "LightOnOCR 3",
        "id": "lightonocr-3",
        "variantId": "0.8b",
        "variantName": "0.8B"
      },
      "inferenceMs": 3809.031500015408,
      "inputKind": "image",
      "loadMs": 4822.185417,
      "maxNewTokens": 1024,
      "memoryMeasurementScope": "Process lifetime peakRSS through complete output construction; not a validated minimum.",
      "modelId": "lightonocr-3-0.8b-native",
      "neuralInferencePerformed": true,
      "outputLimitReached": false,
      "outputTokens": 77,
      "pageInferenceMs": [
        3809.031500015408
      ],
      "pagesPerSecond": 0.2625339275865676,
      "parameterDtype": "torch.float32",
      "pdfScale": 1.5,
      "peakResidentMemoryBytes": 5678432256,
      "placement": {
        "mps:0/torch.float32": 473
      },
      "precision": "fp32",
      "provider": "Metal/MPS",
      "requestedDevice": "mps",
      "requestedPrecision": "fp32",
      "reused": false,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "tokensPerSecond": 20.215112424165703,
      "torchVersion": "2.10.0",
      "transformersVersion": "5.22.1",
      "version": "5.22.1",
      "visualizationMs": 1.939375011716038,
      "totalMs": 8631.216917015408,
      "grade": "interactive"
    }
  }
]

LightOnOCR 3 4B

1 configurations tested
Passed1
Failed0
TestedNov 14, 2026
Runtimetransformers 5.22.1

Quality criteria

Saved paragraph recall, chart location and completed table, receipt text, ordered PDF, EOS and per-request speed limit.

PrecisionDeviceResultPlatformInference / load
FP16mpsAcceleratedPassedNot measuredmacos · aarch645582 ms / 58789 ms
Test setup & sample details
Quality gate
Saved paragraph recall, chart location and completed table, receipt text, ordered PDF, EOS and per-request speed limit.
Tested
2026-11-14T11:47:22.328635+00:00
Environment
Apple M3 Max macOS 27.2; Transformers 5.22.1 and PyTorch 2.10.0
Runtime
transformers 5.22.1
Artifact revision
b71010f095dc3735de7ee6b9969bff8f8f50b39a
Fixture
{
  "file": "model-samples/desktop/lightonocr2/fixtures/scientific-paragraph.png",
  "url": "https://assets.localai.tools/samples/v1/files/de0959e7c3fe59f4d706bb84d993cd1aae000a6b25bcef96f472d3a675871dfb/scientific-paragraph.png",
  "size": 26674,
  "sha1": "b4e9dd3ea6464bc8e1ff1f95134eccfc25622d52",
  "sha256": "de0959e7c3fe59f4d706bb84d993cd1aae000a6b25bcef96f472d3a675871dfb"
}
Rerun when
artifact revision/weight change; pinned runtime/prediction/preprocessing change; fixture/checksum or quality-gate change; relevant host/backend behavior change
Per-case quality & failure details
[
  {
    "precision": "fp16",
    "device": "mps",
    "status": "pass",
    "quality": {
      "passed": true,
      "content": true,
      "eos": true,
      "speed": true,
      "metric": "reference-token-recall",
      "score": 0.95,
      "minimum": 0.9
    },
    "runtime": {
      "acceleration": "partial",
      "accelerationDetails": "Vision/language parameters, inputs and vision activations on selected device; rendering, tokenizer and decoding on CPU",
      "activationPlacement": {
        "languageOutput": "mps:0",
        "visionOutput": "mps:0"
      },
      "allPagesReachedEos": true,
      "artifactRevision": "b71010f095dc3735de7ee6b9969bff8f8f50b39a",
      "cpuThreads": 4,
      "decoding": "greedy",
      "device": "mps",
      "driverAllocatedMemoryBytes": 9549774848,
      "durationMs": 6079.231124999999,
      "engine": "transformers",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "lightonocr-3-native",
        "defaultModelId": "lightonocr-3-0.8b-native",
        "displayName": "LightOnOCR 3",
        "id": "lightonocr-3",
        "variantId": "4b",
        "variantName": "4B"
      },
      "inferenceMs": 5582.385957997758,
      "inputKind": "image",
      "loadMs": 58788.973667,
      "maxNewTokens": 1024,
      "memoryMeasurementScope": "Process lifetime peakRSS through complete output construction; not a validated minimum.",
      "modelId": "lightonocr-3-4b-native",
      "neuralInferencePerformed": true,
      "outputLimitReached": false,
      "outputTokens": 77,
      "pageInferenceMs": [
        5582.385957997758
      ],
      "pagesPerSecond": 0.17913487306754966,
      "parameterDtype": "torch.float16",
      "pdfScale": 1.5,
      "peakResidentMemoryBytes": 18745262080,
      "placement": {
        "mps:0/torch.float16": 723
      },
      "precision": "fp16",
      "provider": "Metal/MPS",
      "requestedDevice": "mps",
      "requestedPrecision": "fp16",
      "reused": false,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "tokensPerSecond": 13.793385226201325,
      "torchVersion": "2.10.0",
      "transformersVersion": "5.22.1",
      "version": "5.22.1",
      "visualizationMs": 1.4462909894064069,
      "totalMs": 64371.359624997756,
      "grade": "interactive"
    }
  }
]
Release validation still required
  • Publication, public backup and dependent recovery are deferred.
  • Windows verification is deferred.

Something not quite right?

Tell us what happened with LightOnOCR 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.