What is Chandra OCR 2?
Chandra OCR 2 is a document-parsing model from Datalab, the company behind Marker and Surya, released on 18 March 2026 with version 0.2.0 of the chandra-ocr package. Datalab calls it a 4B model; the Hugging Face safetensors count is 5.30 billion parameters, roughly 10.6 GB in BF16. It is built on Qwen 3.5 and turns page images and PDFs into Markdown, HTML or JSON with layout blocks and bounding boxes.
The model labels 18 block types, from tables and equations to chemical diagrams and tables of contents, converts flowcharts to Mermaid, pulls values out of charts and reads handwriting and filled-in forms with checkboxes. Datalab claims support for 90 or more languages.
The code is Apache 2.0. The weights use a modified OpenRAIL-M licence: free for research and personal use, and free commercially only if your company has both under $2M of gross revenue in the prior year and under $2M of total funding. Separately, no company of any size may use it in a product that competes with Datalab's own. Above the thresholds a paid licence is needed; the price is not published.
How much does Chandra OCR 2 cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free under $2M | Modified OpenRAIL-M; no competing use; paid licence above the threshold |
| Datalab Convert API | $4 per 1,000 pages | Fast or balanced mode; accurate $10 |
| Team | $400/month | Includes $400 of usage |
Weights free for research, personal use and companies under $2M revenue and $2M funding; paid licence above that (price not published). Datalab API: Convert $4 per 1,000 pages fast/balanced, $10 accurate; Team plan $400/month.
Is Chandra OCR 2 worth it?
Chandra OCR 2 is one of the few OCR models whose accuracy claims have held up under someone else's test. On the BHL OCR leaderboard, an independent board of 2,165 hand-transcribed pages from 18th- and 19th-century natural-history books, it ranks 2nd of 26 with a character error rate of 0.0214, statistically tied with dots.mocr at 0.0214 and dots.ocr at 0.0224. Its loop rate there, 0.09% or two pages, is the lowest of any generative model on the board.
Datalab's own figures point the same way. It reports 85.8 on Ai2's olmOCR-Bench, the top open-weights score in its table, ahead of dots.ocr 1.5 at 83.9 and olmOCR 2 at 82.4; the release notes say 85.9, so expect a decimal of drift. Jina and OpenBMB, both rivals, list Chandra 2 at 85.8 in their own olmOCR tables. On Datalab's internal 43-language set it averages 77.8%, against 67.6% for Gemini 2.5 Flash, but it still trails Gemini on Arabic, Hindi, Urdu and Vietnamese.
Size is the cost. At 5.30 billion parameters by Hugging Face's count, it is about five times PaddleOCR-VL-1.6, and Datalab's throughput figure is 1.44 pages a second on one H100 with 96 concurrent requests on the hard olmOCR set, which it estimates at about 2 pages a second on ordinary documents. Surya OCR 2.5, from the same company, manages 4.52 pages a second. The BHL board also found its diplomatic error rate, which keeps long-s and ligatures as printed, at 0.093, much worse than its headline score, partly because of the markup in its output.
The licence decides most purchases. The $2M ceiling applies to revenue and funding separately, so a startup that has raised a $3M seed round already needs a paid licence, and the clause banning use in any product that competes with Datalab's API catches document-processing vendors of every size. Surya OCR 2.5 lifts the threshold to $5M; PaddleOCR-VL-1.6, GLM-OCR and MiniCPM-OCR 1.0 have no threshold at all.
Where it earns its place
A digitisation team at a natural-history library
Transcribe 19th-century printed volumes for full-text search
Serve the model with chandra_vllm and batch page images through the chandra CLI. Research and non-profit use is free under the licence, and the BHL board's 0.0214 error rate on this kind of material is a measured figure rather than a vendor claim Keep a line-level tool alongside it if you need original spellings and ligatures.
An engineer at a seed-stage insurtech with $1.5M raised
Extract handwritten claim forms, including checkboxes, into JSON
Self-host on one H100 or a rented equivalent; at Datalab's estimate of 2 pages a second that is about 7,200 pages an hour. The weights are free while both revenue and funding stay under $2M, so the next funding round triggers a licence negotiation.
An operations lead at a 200-person logistics firm above the threshold
Parse supplier PDFs without buying a model licence
Send pages to Datalab's hosted Convert API at $4 per 1,000 pages in fast or balanced mode or $10 in accurate mode, so 50,000 pages a month costs $200 to $500. Datalab's card says the hosted API scores 86.7 on olmOCR-Bench, slightly above the open weights.
Before you commit
- Setup
- pip install chandra-ocr, start a vLLM server with chandra_vllm, then run chandra input.pdf ./output; a Transformers route needs pip install chandra-ocr[hf] and torch.
- Integrations
- vLLM (recommended) and Hugging Face Transformers; Python InferenceManager API; Markdown, HTML and JSON output with bounding boxes; Datalab hosted API and playground.
- Data and privacy
- Code Apache 2.0; weights modified OpenRAIL-M with a $2M revenue and $2M funding ceiling and a no-competing-use clause. Self-hosted runs keep documents on your hardware.
- Cost at scale
- Weights about 10.6 GB in BF16; Datalab measures 1.44 pages a second on one H100 80GB at 96 concurrent sequences on olmOCR-Bench pages and estimates about 2 on real-world documents. Hosted: $4 per 1,000 pages fast, $10 accurate.
- Limits
- P95 latency of 156 seconds in Datalab's own throughput test; weaker than Gemini 2.5 Flash on Arabic, Hindi and Urdu in Datalab's table; the vendor client resamples looping pages, which adds latency.
What’s good
- 2nd of 26 on the independent BHL OCR board, tied with the leader
- 85.8 on olmOCR-Bench, top open-weights score in Datalab's table
- Layout blocks, handwriting, forms, charts and Mermaid diagrams in one model
- 0.09% loop rate in independent testing
What’s not
- Free commercial use only under $2M revenue and $2M funding
- Bans use in products that compete with Datalab's API
- 5.30B parameters; about 1.44 pages a second on an H100 on hard pages
- Diplomatic (as-printed) error rate of 0.093 on the BHL board, far above its headline
Best for
Researchers, archives and small companies under the licence threshold who want the most independently tested open parser for difficult scans and handwriting.
Chandra OCR 2 alternatives
Chandra OCR 2: common questions
- Is Chandra OCR 2 worth it?
- Chandra OCR 2 has the best outside evidence of any open document parser we cover, and the tightest licence: free commercial use ends at $2M. We rate it 4.2 out of 5.
- How much does Chandra OCR 2 cost?
- Weights free for research, personal use and companies under $2M revenue and $2M funding; paid licence above that (price not published). Datalab API: Convert $4 per 1,000 pages fast/balanced, $10 accurate; Team plan $400/month.
- What are the downsides of Chandra OCR 2?
- Free commercial use only under $2M revenue and $2M funding; bans use in products that compete with Datalab's API; 5.30B parameters; about 1.44 pages a second on an H100 on hard pages; diplomatic (as-printed) error rate of 0.093 on the BHL board, far above its headline.
- Who is Chandra OCR 2 best for?
- Researchers, archives and small companies under the licence threshold who want the most independently tested open parser for difficult scans and handwriting.
- How much does Chandra OCR 2 cost for a team or in production?
- Weights about 10.6 GB in BF16; Datalab measures 1.44 pages a second on one H100 80GB at 96 concurrent sequences on olmOCR-Bench pages and estimates about 2 on real-world documents. Hosted: $4 per 1,000 pages fast, $10 accurate.
- Is Chandra OCR 2 safe for company data?
- Code Apache 2.0; weights modified OpenRAIL-M with a $2M revenue and $2M funding ceiling and a no-competing-use clause. Self-hosted runs keep documents on your hardware.
- What are the best Chandra OCR 2 alternatives?
- The closest alternatives we have reviewed are Surya OCR 2.5, dots.ocr, LightOnOCR-3-4B, PaddleOCR-VL-1.6 and Mistral OCR 4.1.
Answers checked Nov 2, 2026.
Make Chandra OCR 2? Show this rating on your site
