Nanbeige4.2-3B vs Baidu Unlimited-OCR-2

Nanbeige4.2-3B 85

Memory 30, the smallest build from Nanbeige is the FP8 repository (the 8-bit GPTQ one is the same size), 5.19 GB, and 5.19 GB × 1.2 = 6.2 GB, within 16 GB. bartowski's GGUF isn't read for this line. Runtimes 15, Released SGLang 0.5.23 has the model (srt/models/nanbeige.py) with its parsers, llama.cpp b12242 has the nanbeige architecture, and vLLM 0.33.0 names NanbeigeForCausalLM in its registry, served through its Transformers backend. Builds 15, Quantised builds from Nanbeige at two precisions, FP8 and 8-bit GPTQ, each in its own repository. Terms 15, Apache 2.0 in the card's metadata, an OSI licence, with no use policy or other terms on the card. Download 10, the Hub API reports gated false for the model and its FP8 and 8-bit repositories on 10 November 2026, so the files download with no account and nothing to accept. Hosted 0, Hugging Face maps no inference provider to it, and no Nanbeige or BOSS Zhipin API for it was found. 30 + 15 + 15 + 15 + 10 + 0 = 85.

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.

Nanbeige4.2-3B 72

Model card 17, Intended use, limitations, the architecture, the parameters (3B outside the embeddings), the context and the languages (in the metadata) are on the card. Training data 5, A general description of how the fine-tuning and reinforcement learning data was built and filtered, with no sources, sizes or cut-off. Last change 15, the last commit, on 11 September 2026, updated modeling_nanbeige.py, 60 days before this check and within 90. Responsiveness 20, Of the 34 most recent discussions that aren't pull requests (21 July to 10 September 2026), 20 have a reply from a Nanbeige member, which is most. Version policy 5, Nanbeige's earlier models (Nanbeige4-3B, 4.1-3B) stay in their own repositories, as the Base, FP8, 8-bit and DSpark builds of this one do. File health 10, Config, tokenizer config, template and card agree on the architecture and the 262,144 tokens. 17 + 5 + 15 + 20 + 5 + 10 = 72.

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).

Nanbeige4.2-3B 55

File safety 15, safetensors only, no .bin, .pt or .pth files (25). Provenance 15, the files sit in the Nanbeige organisation (Nanbeige LLM Lab), which the lab's arXiv report for this model links, so the publisher's own account, although the Hub hasn't verified it (10). Safety evaluation 0, no safety results on the card. Agent guidance 5, General cautions: the limitations section says outputs may be inaccurate, biased or harmful. Licence risk 15, Apache 2.0 has no use policy, no threshold and no clause letting the licensor change or revoke it. Security programme 5, BOSS Zhipin runs a security response centre for reporting flaws in its products, as read for the Nanbeige4.3-33B listing (5), though nothing links it to the model. 15 + 15 + 0 + 5 + 15 + 5 = 55.

Baidu Unlimited-OCR-2 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.

Nanbeige4.2-3B 85

Chat template 20, A Jinja chat template in tokenizer_config.json (7,961 characters), with enable_thinking, preserve_thinking, tools and tool_call_format switches. Tool-calling format 25, the card says passing tools switches on the tool template and recommends tool_call_format xml, with json for compatibility, and its SGLang and vLLM commands name the nanbeige tool-call parser, which released SGLang 0.5.23 has. Trained tool calling 20, the card describes fine-tuning on tool-use environments and agent scaffolds and reinforcement learning on agent tasks, and the template tells the model it may call one or more functions and renders every call in a turn. Structured output 0, Neither the card nor the files mention JSON or JSON Schema output, beyond json as a tool-call format. Reasoning control 10, the card documents enable_thinking, on by default and False per request, and preserve_thinking for earlier turns, and names the nanbeige reasoning parser, which returns the reasoning apart from the answer. Context 10, 262,144 tokens, which the card says the model supports and config.json gives as max_position_embeddings. 20 + 25 + 20 + 0 + 10 + 10 = 85.

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.