All tasks / Models
LightOnOCR-3 (4B, 1B, 0.8B)
LightOn's open-weights end-to-end OCR model family that turns page images into Markdown with layout grounding.
- LightOn
- Open weights
- Released 6 Oct 2026
Where it ranks
- –Documents & OCRModels · Local · no result on this task’s benchmarks yet
Facts
- Vendor
- LightOn
- Family
- LightOnOCR
- Released
- 6 Oct 2026
- Weights
- Open: download and run it yourself
- License
- Apache-2.0
- Parameters
- 4.54B
- GPU memory to run locally
- ~11 GB (Estimate, no official figure: BF16 weights ≈ 4.54 B × 2 GB + ~1.5 GB runtime overhead. Size is the Hugging Face safetensors total.)
- Access
- Free (open weights)
- Last verified
- 19 Oct 2026
Price from huggingface.co, 19 Oct 2026.
Benchmark results
| Benchmark | Result | Entry on the leaderboard | As of | Source |
|---|---|---|---|---|
| ParseBench (models) | 75.1% | LightOnOCR-3-4BOverall (mean of tables, charts, content faithfulness, semantic formatting, visual grounding) from leaderboard.csv in the ParseBench repo (last leaderboard commit 2026-10-09). Best of three sizes listed (1B: 71.4, 0.8B: 74.6). | 9 Oct 2026 | github.com |
Where you can use it
No tracked tool lists this model yet; as open weights it runs in local runtimes.
Earlier versions
- LightOnOCR-2-1B · 16 Jan 2026
- LightOnOCR-1B-1025 · 20 Oct 2025