Chandra OCR 2
Datalab's OCR model for 90+ languages with Markdown, HTML or JSON output and layout. Free for research and small teams, not for rivals.
In short. Chandra OCR 2 turns scans, PDFs and photos into Markdown, HTML or JSON and keeps the layout, including tables, forms with checkboxes, math and handwriting. It supports more than 90 languages. It is one of the most downloaded OCR models, but its license is not free for larger companies.
Summary
Datalab, the US company behind the Marker and Surya tools, announced Chandra OCR 2 on 18 March 2026. The blog calls it a 4B model, down from 9B for Chandra 1; the BF16 checkpoint holds 5.3 billion parameters and is about 10.6 GB, built on the Qwen3.5 vision-language architecture. It outputs Markdown, HTML or JSON with more than 15 types of layout blocks and bounding boxes, writes Mermaid code for diagrams, returns structured data for charts, and captions images. It handles handwriting and rebuilds forms, including checkboxes. Datalab says it supports more than 90 languages and reaches 77.8% on its internal multilingual test over the top 43.
Datalab reports 85.8% on olmOCR-Bench, which it calls state of the art; its launch post said 85.9%, and the model card was later corrected to 85.8 after normalizing the scoring. Its table lists Chandra 1 at 83.1, olmOCR 2 at 82.4 and DeepSeek OCR at 75.4. The blog gives about 2 pages per second on an H100 with 96 concurrent requests. vLLM is the recommended way to run it, through the chandra-ocr package. A hosted Datalab API, which runs an improved version, scores 86.7. All of these numbers are Datalab's own.
The license is a modified Open RAIL-M, not Apache or MIT. It is free for research, personal use, and organizations with under US$2 million of revenue and under US$2 million raised. Larger organizations need a commercial license, and organizations that offer a product or service competing with Datalab's cannot use the model under this license. Derivatives and even outputs must carry the same license, which matters if you plan to sell a service built on it.
Good for
- Research, personal projects and early-stage startups that need strong open OCR in many languages
- Handwriting, forms with checkboxes, math and complex tables
- Getting layout blocks with bounding boxes and Markdown, HTML or JSON in one pass
- Teams that may later buy the hosted API or a commercial license
Watch out for
- Commercial use is restricted above US$2 million in revenue or funding, and the license bars competing products or services
- Share-alike applies to outputs and derivatives, which is unusual and needs a legal read
- The 4B label in the blog and the 5.3B parameter count in the checkpoint differ, so size your memory from the file
- Benchmark numbers are Datalab's own, and the launch score was later corrected from 85.9 to 85.8
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the Chandra OCR 2 variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| Chandra OCR 2 BF16 | 5.3B | 12 GB | 6.09 GB | 3.66 GB | Laptop or phone |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
A modified Open RAIL-M license. Free for research and personal use, and for commercial use only by organizations with under US$2 million in prior-year revenue and under US$2 million raised in total. It also bans using the model to compete with Datalab and requires derivatives and outputs to carry the same license.
- Attribution or model name must be shown
- Free commercial use only below a revenue cap
- Acceptable-use policy applies
More about OpenRAIL-M (Datalab)
See also
- Best open-source OCR and document models in 2026
- Free until you grow: where open licenses draw the line
- Eleven open OCR models in a year: small and hard to compare
Frequently asked questions
- What is Chandra OCR 2?
- Chandra OCR 2 turns scans, PDFs and photos into Markdown, HTML or JSON and keeps the layout, including tables, forms with checkboxes, math and handwriting. It supports more than 90 languages. It is one of the most downloaded OCR models, but its license is not free for larger companies.
- Can I use Chandra OCR 2 commercially?
- Yes, with conditions. Chandra OCR 2 is released under the OpenRAIL-M (Datalab), which allows commercial use but adds extra rules: Attribution or model name must be shown; Free commercial use only below a revenue cap; Acceptable-use policy applies. Read the license text before you ship a product.
- Is Chandra OCR 2 open source?
- Chandra OCR 2 is open-weight, not open source in the strict sense. You can download the model files, but the OpenRAIL-M (Datalab) adds conditions that the Open Source Initiative definition does not allow.
- Can I run Chandra OCR 2 on my own computer?
- Yes. The smallest variant, Chandra OCR 2, needs about 3.66 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Chandra OCR 2 released?
- Chandra OCR 2 was released by Datalab on March 18, 2026.
- How long a context window does Chandra OCR 2 have?
- Chandra OCR 2 supports a context window of up to 256K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Announcing Chandra OCR 2: 90+ Languages, Top Benchmarks (Datalab, 18 Mar 2026) datalab.to
- Chandra OCR 2 model card (Hugging Face) huggingface.co
- Chandra OCR 2 license (modified Open RAIL-M) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Surya OCR 2 Datalab | 690M | 256K | OpenRAIL-M (Datalab Surya) (Conditional) | 0.476 GB Laptop or phone | |
| LightOnOCR-3 LightOn | 853M – 4.54B | 256K | Apache 2.0 (Permissive) | 0.589 GB Laptop or phone | |
| PaddleOCR-VL Baidu | 960M | – | Apache 2.0 (Permissive) | 0.662 GB Laptop or phone | |
| DeepSeek-OCR 2 DeepSeek | 3.39B | – | Apache 2.0 (Permissive) | 2.34 GB Laptop or phone |