Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.
Analysis8
github.comauthorityabout the chandra family1,927 words
Chandra 2 achieved an overall accuracy of 85.8 ± 0.8 on internal benchmarks while supporting 90+ languages. The model processes documents at an estimated 2 pages per second in real-world usage on a single NVIDIA H100 80GB GPU.generated, cited
nullkdev.github.io2026-07-29about the chandra family1,128 words
Chandra 2 is a 5-billion-parameter model built on Qwen 3.5 that achieves 85.9% on the olmOCR benchmark and processes 1.44 pages/second on an H100. The weights ship under a modified OpenRAIL-M licence allowing free use for startups under $2M in funding or revenue but prohibiting competition with Datalab's API. RealDocBench indicates that models scoring well on standard benchmarks often show significantly lower accuracy on authentic regulatory documents with field-level requirements.written, reviewed, cited
- The multilingual benchmark is Chandra's own exam where the authors state no good public multilingual alternative exists.
- A page that is 95% correctly transcribed can be 100% wrong about the one field you needed.
- The weights ship under a modified OpenRAIL-M licence that explicitly cannot be used to compete with Datalab's own API.
datalab.to1,553 words
Datalab released Chandra 2, a 4B parameter OCR model that achieves 85.9% on the olmOCR benchmark. The model averages 77.8% on the top 43 languages and processes 2 pages per second on an H100 with 96 concurrent requests. It outputs 15+ block types with bounding boxes and handles complex layouts including tables, math, and Mermaid diagrams.written, reviewed, cited
- Chandra 2 supports 90 world languages, including significant gains in Indic scripts like Kannada and Malayalam.
- The model includes an MTP head to further improve latency for on-prem deployment scenarios.
- The hosted API scores 86.7% on olmOCR overall, outperforming the open model's 85.9%.
beri.net1,142 words
Chandra OCR 2 is a 5B-parameter vision-language model fine-tuned from Qwen 3.5 that converts PDFs and images into structured Markdown, HTML, or JSON while preserving layout. Released in March 2026, it currently tops olmOCR-Bench with an overall score of 85.8, significantly outperforming olmOCR's 82.4 on complex tables and long tiny text. The code is Apache 2.0, but the weights use modified OpenRAIL-M licensing, allowing free use for research and startups under $2M in funding or revenue, while commercial self-hosting requires a paid license.written, reviewed, cited
- The hosted Datalab API provides zero data retention and SOC 2 Type 2 compliance for sensitive document processing.
- The vLLM backend achieves approximately 1.44 pages per second on a single H100 80GB GPU at 96 concurrent sequences.
pypi.orgabout the chandra family2,575 words
Chandra OCR 2 converts images and PDFs into structured HTML, Markdown, or JSON while preserving layout information. Released in 3/2026, the model supports 90+ languages and offers two inference modes: local (HuggingFace) and remote (vLLM server). The code uses an Apache 2.0 license while model weights require a modified OpenRAIL-M license.written, reviewed, cited
- Engineers can start their own vLLM server using the datalab-to/chandra-ocr-2 model instead of relying on the hosted API.
- The model weights license prohibits competitive use with the company's hosted API unless the user obtains broader commercial licensing.
mlhive.com2026-04-181,633 words
Chandra OCR 2 is a 5-billion parameter open-source model that processes documents as direct image-to-markdown translations. The team processed 27,000 arXiv papers in under 30 hours, sustaining a speed of approximately 3.5 pages per second. Deploying on an 8x H100 cluster costs roughly $900 for the full test, utilizing torch.compile and Flash Attention 2 for efficiency.written, reviewed, cited
- Models in the 1B range frequently hallucinate numbers in dense financial tables or drop crucial negative signs in mathematical formulas.
newreleases.io2026-03-18about the chandra family307 words
