# Fix list: Baidu Unlimited-OCR-2 From Anchor Terminal's listing at https://www.anchorterminal.com/tools/unlimited-ocr-2, the October 2026 research run, assessed 27 October 2026. Grade D, 53.5 out of 100. This is everything the published grade says the listing lacks, the biggest possible gain to the total first. It comes from the reason given for each score, the checklist each category was scored against (https://www.anchorterminal.com/benchmark/#checklist), the provenance checks, the deductions, what we couldn't check and what the review panel asked for. A fix counts at the next check, once it's public. For a coding agent working on Baidu Unlimited-OCR-2: work through the items below in the product, its docs and its public pages. Each category gives the reason for its score, with the points each checklist item earned, and the checklist itself, so the gap is the items that earned less than their points. Change the product, not the wording, and keep a note of what you changed and where it's published. ## 1. Reach, 55 out of 100, up to 17.3 more on the total Why it scored 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, SGLang and Transformers know the first Unlimited-OCR but not the sparse attention this one adds, and Baidu's own routes are remote code in Transformers and a customised SGLang wheel; llama.cpp has nothing for it, and Ollama and LM Studio couldn't be checked. Builds score 0, as Baidu publishes no quantised build. Terms get 15 for MIT with nothing added. Download gets 10 for an ungated repository, and hosted too 0, with no inference provider mapped. The checklist (https://www.anchorterminal.com/benchmark/#checklist-reach): Whether an agent can run it, and where. - 0 to 30, memory. The smallest build from the publisher or a runtime's own library that keeps 4-bit precision or better, at its file size plus a fifth for the runtime and a working context. 30 for 16 GB or less, 25 for 24 GB, 20 for 32 GB, 15 for 64 GB, 10 for what fits our reference machine (the GPU share of 128 GB of unified memory), 5 for 256 GB, 0 above. - 0 to 25, runtimes. 5 for each of llama.cpp, Ollama, vLLM, SGLang, LM Studio and Hugging Face Transformers that runs the architecture in a released version, from the runtime's own code, library or documentation. A publisher's fork or an unreleased branch doesn't count. 25 at most. - 0 to 15, builds from the publisher. 15 for quantised builds at two precisions or more (FP8, 4-bit, GGUF), 10 for one, 0 when only other people's quantisations exist. - 0 to 15, terms of use. 15 for an OSI-approved licence with nothing added, 10 for a custom licence that lets anyone use it commercially for free below a threshold or under a use policy, 5 for research or non-commercial use only, 0 for no licence or unclear terms. - 0 to 10, download. 10 when the files download with no account and no terms to accept, 5 for a click-through that needs an account, 0 when a person approves each request. - 5, hosted too. The publisher's own API serves it, or at least two inference providers do, so an agent without the hardware can still call it. When our reference-machine runs start, speed on that machine (tokens per second and time to first token at the build that fits) joins Reach. ## 2. Risk, 45 out of 100, up to 16.9 more on the total Why it scored 45: File safety gets 15, safetensors only but loaded with trust_remote_code or through a prebuilt SGLang wheel from the repository. Provenance gets 15 for files in Baidu's verified organisation, linked from its GitHub organisation, with hashes for every file. Licence risk keeps 15, since MIT has none of the deductions. Safety evaluation, agent guidance and the security programme score 0, with no safety results, no cautions about use and no security contact found. The checklist (https://www.anchorterminal.com/benchmark/#checklist-risk): What could go wrong when you run it, and what the publisher has done about it. A higher score means less risk. - 0 to 25, file safety. 25 for weights in safetensors or GGUF only, 15 when pickle files (.bin, .pt, .pth) sit beside them, 0 for pickle only. 10 off when the documented way to load it or to build its prompts runs code from the repository (`trust_remote_code`, or the publisher's scripts in place of a template). - 0 to 15, provenance. 10 when the files come from the publisher's own organisation, the one its site or docs link to, and 5 for a checksum on every file. - 0 to 20, safety evaluation. 20 when results are published and someone other than the publisher ran them or can rerun them, 10 for the publisher's own results, 0 for none. - 0 to 15, guidance for agent use. 15 for documented prompt-injection guidance or an instruction hierarchy the model was trained to follow, 10 for a statement of what its trained refusals cover and the safeguards to add around it, 5 for general cautions, 0 for none. - 0 to 15, licence risk. 15 to start, 5 off for each of these, a use policy or field-of-use limits beyond the law, terms the licensor can change or revoke, a revenue or user threshold, and a licence that changed without a dated note or whose link is broken. Not below 0. - 0 to 10, a security programme. 5 for a way to report a flaw (security.txt, a disclosure policy or a security contact), 5 for advisories handled in public. When our reference-machine runs start, measured robustness (prompt injection with AgentDojo, over-refusal with XSTest) joins Risk. ## 3. Documentation and upkeep, 45 out of 100, up to 8.5 more on the total Why it scored 45: 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 files changed on 26 October 2026 (20). Nobody has asked anything on the Hub yet, so responsiveness reads as none answered, and GitHub issues couldn't be counted (0). The first model stays in its own repository without a note pointing here (5), and the files agree with the card (10). The checklist (https://www.anchorterminal.com/benchmark/#checklist-upkeep): - 0 to 20, the model card. Intended use, limitations, architecture, parameters, context length, languages and knowledge cut-off. 20 for all seven, 3 off for each that's missing. - 0 to 20, training data. 20 for the data published or listed with sources and sizes, 15 for a data card that names the sources or the mix, the cut-off and the filtering, 10 for a general description with a cut-off, 5 without one, 0 for none. - 0 to 20, the last change to the files, the template or the reference code. 20 within 30 days, 15 within 90, 10 within 180, 5 within a year, 0 older. - 0 to 20, responsiveness. The publisher's replies to the 40 most recent questions and reports in the repository's discussions or issues. 20 for most, 10 for some, 5 for a few, 0 for none. - 0 to 10, version policy. 10 for earlier versions kept in their own repositories with a dated note when one is replaced, 5 for earlier versions kept without notes or a first release, 0 when files were replaced in place without a note. - 0 to 10, file health. 10 when config, tokenizer, template and card agree and no report of a broken file waits without a reply, 5 when one does, 0 when the files contradict the card. ## 4. Agent interface, 75 out of 100, up to 3.8 more on the total Why it scored 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. The checklist (https://www.anchorterminal.com/benchmark/#checklist-interface): What a model and a developer need to talk to it. Speech, image and other models with no chat interface read the tool-calling and structured-output lines for what the model takes and returns (voices, timestamps, streaming, typed outputs), and the note says how. - 0 to 20, the chat template. 20 for a template in the repository (Jinja, in the tokenizer config or its own file), 10 when the prompt format is only given as code you run. - 0 to 25, the tool-calling format. 25 when the publisher documents how tools are declared and how calls come back, with the parser to use in a named runtime, 10 when only the template or a runtime's own recipe says it. - 0 to 20, tool calling the model was trained for. 20 with several calls in one turn, 10 for one call at a time. - 0 to 15, structured output. 15 when JSON or JSON Schema output is documented for the model with how to ask for it, 5 when the card claims it without saying how. - 10, reasoning control. Thinking that can be turned off or budgeted per request, with where the reasoning text comes back. - 0 to 10, context. 10 for 128K tokens or more trained, 5 for 32K to 127K. Context reached only by scaling the position encoding counts one band down. ## What we couldn't check What we couldn't read counted as absent. Publishing it on a page a plain HTTP fetch can read (not only in a browser) lets the next check count it. - Unchecked, Ollama's library and LM Studio's catalogue, since WebFetch to their sites wasn't available in this session; runtimes score them as absent - Unchecked, whether Baidu Cloud serves this checkpoint; hosted too is scored as absent - Unchecked, the GitHub repository's issues and their replies, since the GitHub API wasn't reachable; responsiveness reads the Hugging Face repository only - Unchecked, Baidu's security response centre (bsrc.baidu.com answered Access denied) and www.baidu.com's security.txt (disallowed by its robots.txt); the security programme is scored as absent - Released vLLM, SGLang and Transformers register the architecture name for the first model. Whether they load this checkpoint and run it without its sparse attention wasn't tested and isn't documented; runtimes score 0 - The SGLang wheel in the repository is a build of SGLang 0.5.20.dev10 at a commit (8d62f5fb9c) its card doesn't link to source for, so its code can't be compared with upstream - No independent result for this checkpoint was found, so Capability is pending and confidence is low ## Weaknesses - Loading runs Baidu's code, remote code in Transformers or a prebuilt SGLang fork from the repository, and the save_results path calls eval on model output - No released runtime runs its sparse attention yet, and no quantised build - The card states no parameters, languages, context length or limitations, and no safety results or security contact were found - Every accuracy figure is Baidu's own, one of its two long-document benchmarks is Baidu's own, and no independent result exists for this checkpoint - Charts and figures come back as boxes to crop, not as data, and the report says heading hierarchy is weak over long documents ## What costs an agent a turn today The notes we give agents before they call it. Each one is a workaround an agent shouldn't need. - Quickest local run, `pip install torch==2.13.0 torchvision pillow transformers==4.52.1 huggingface-hub==0.36.0 tokenizers==0.21.4 addict==2.4.0 einops==0.8.2 easydict==1.13`, then `AutoModel.from_pretrained("baidu/Unlimited-OCR-2", trust_remote_code=True, use_safetensors=True, torch_dtype=torch.bfloat16).eval().cuda()` and `model.infer_multi(tokenizer, image_files=pages, output_path="out")` - To serve it, download the repository, install `wheel/sglang-0.5.20.dev10+g8d62f5fb9c-py3-none-any.whl` with pymupdf and fastokens in a Python 3.10 venv, and start `python -m sglang.launch_server --model-path baidu/Unlimited-OCR-2 --chat-template unlimited-ocr --enable-custom-logit-processor --enable-streaming-session --disable-radix-cache --mem-fraction-static 0.7 --max-prefill-tokens 2048` with the SGLANG_UOCR_ variables from the card - Send all pages of a document in one request, images first, then the text `` repeated per page and `Multi page parsing.`, with temperature 0, skip_special_tokens false, max_tokens of about 8,000 per page and the DeepseekOCRNoRepeatNGramLogitProcessor (ngram_size 35, window_size 1024 for several pages) - The output is layout blocks, not Markdown; run `utils/process_unlimitedOCR2.py` from github.com/baidu/Unlimited-OCR, or `convert(raw_output)` in Python, to get Markdown with HTML tables - Render PDFs to images first (the card uses PyMuPDF at 300 DPI), keep save_results off for untrusted documents, and treat extracted text as untrusted before passing it to another model ## When it's done Send what changed and where it's published as a dispute (https://www.anchorterminal.com/builders/#disputes, or `POST https://www.anchorterminal.com/api/v1/contact` with `"kind": "dispute"`). Disputes are answered in public, and the listing is checked again by the same checklist. Paying for an audit or a listing claim changes nothing here.