Surya OCR 2.5 vs Baidu Unlimited-OCR-2

Surya OCR 2.5 90

Memory 30, Datalab's own Q8_0 GGUF is 0.63 GB with a 0.21 GB projector, and 0.84 GB × 1.2 = 1.0 GB, within 16 GB. The BF16 safetensors file is 1.38 GB. Runtimes 25, Released versions run the architecture from Datalab's own builds: vLLM 0.33.0, SGLang 0.5.23 and Transformers 5.22.2 from the safetensors file, llama.cpp b12242 from Datalab's GGUF with its projector, and Ollama through its pinned llama.cpp (b12177 at v0.40.16). Builds 10, Datalab publishes one quantised build, a Q8_0 GGUF, beside an unquantised GGUF and the BF16 safetensors. Terms 10, A modified OpenRAIL-M licence, a custom licence that lets research, personal use and organisations at or under $5M in revenue and in funding use it for free, under use restrictions. Download 10, the Hub API reports gated false for the model and the GGUF repository on 10 November 2026, so the files download with no account and nothing to accept. Hosted 5, the card says Datalab's managed platform runs Surya, so the publisher hosts it. 30 + 25 + 10 + 10 + 10 + 5 = 90.

Baidu Unlimited-OCR-2 55

Memory earns the full 30, since the only build is 6.69 GB of BF16 weights that need about 8 GB with the fifth added. Runtimes score 0. Released vLLM 0.33.0, SGLang 0.5.23 and Transformers 5.22.1 know the first Unlimited-OCR but not the sparse attention this one adds, Baidu's own routes are remote code in Transformers and a customised SGLang wheel, llama.cpp b12111 has nothing for it, so Ollama has nothing either, and LM Studio couldn't be checked. Builds score 0, as Baidu publishes no quantised build (16 exist, all by others). Terms get 15 for MIT with nothing added. Download gets 10 for an ungated repository, and hosted too 0, with no inference provider mapped and the card's Baidu Cloud line unconfirmed.

Surya OCR 2.5 57

Model card 17, Intended use, limitations, the architecture (a Qwen 3.5-style vision-language model with separate detection models), parameters (689M), the serving length (24,576 tokens) and 91 languages are on the card. Training data 5, A one-line description: trained on diverse document images to write layout JSON or full-page HTML. No sources, sizes or cut-off. Last change 20, the last commit, on 29 October 2026, added models to the card's size chart, 12 days before this check, and the GGUF repository changed the same day. Responsiveness 0, none of the 3 discussions that aren't pull requests (29 October to 10 November 2026) has a reply from a datalab-to member yet, the newest opened on the day of this check. Version policy 5, Surya OCR 2 stays in its own repository and 2.5 came out in a new one. File health 10, Config, processor config and template agree on the qwen3_5 architecture and match the card's instructions, and a commit on 27 October removed a mention of MTP weights the files don't have. 17 + 5 + 20 + 0 + 5 + 10 = 57.

Baidu Unlimited-OCR-2 50

The card covers intended use and the new attention, and misses limitations, parameters, context length, languages and a cut-off (5). Training data is described by mix and labelling with no datasets named and no cut-off (5). The card changed on 2 November 2026 and the SGLang wheel on 29 October (20). Baidu's uploader replied in 1 of the 5 Hub discussions, a few (5). The first model stays in its own repository without a note pointing here (5), and the files agree with the card (10).

Surya OCR 2.5 70

Chat template 20, chat_template.jinja in the repository (2,872 characters). Tool-calling format 25, Read for a document parser as what it takes and returns. Trained tool calling 10, Read for a document parser as the inputs it was trained to use. Structured output 15, Read for a document parser as typed outputs. Reasoning control 0, A document parser with no reasoning. Context 0, Datalab documents a 24,576-token model length for serving, below the 32K band. 20 + 25 + 10 + 15 + 0 + 0 = 70.

Baidu Unlimited-OCR-2 75

Read for a document parser, the tool-calling, trained tool calling and structured-output lines stand for what it takes and returns. Baidu documents the page images, the prompt, the request settings and the block output, and ships a converter for the server's output (25). It was trained to read many pages in one request (20). The block format with classes, boxes and HTML tables, and its conversion to Markdown, are documented with how to get them (15). The prompt is built in code, with no Jinja template in the repository (10). Reasoning control scores 0, and the context 5, trained at up to 64K tokens.