LLM Models

Jina AI's 0.6B-parameter multilingual listwise document reranker, released 27 July 2026 as a drop-in upgrade to jina-reranker-v3 with an unchanged request schema. It keeps the last-but-not-late (LBNL) listwise interface, ranking a query against a whole candidate list in a single forward pass, and is built on Qwen3-0.6B with 28 layers of hybrid 3L2G attention (sliding-window layers of width 1024 interleaved with global-attention layers and a pinned terminal global layer) plus a lightweight 1024-512-512 MLP projector. It is trained with self-distillation for domain robustness and, per the vendor, runs 1.22x-1.56x faster than v3 on short to long contexts. Weights are CC-BY-NC-4.0.

Fifth-generation multilingual text embedding model from Jina AI, released 18 February 2026. A 677M-parameter model built on the Qwen3-0.6B-Base backbone, trained in two stages: embedding distillation from Qwen3-Embedding-4B, then four task-specific LoRA adapters (retrieval, text-matching, clustering, classification) fitted on frozen backbone weights. It produces 1024-dimensional embeddings via last-token pooling with Matryoshka truncation down to 32 dimensions, handles sequences up to 32,768 tokens through RoPE with adjusted base frequencies, supports asymmetric retrieval with Query:/Document: prefixes, and is regularised for binary-quantisation robustness. Weights are CC-BY-NC-4.0.

Jina AI's multimodal embedding model, released 7 May 2026. A ~1.74B-parameter model that encodes text, images, video, audio and PDF pages into a single shared 1024-dimensional vector space aligned with jina-embeddings-v5-text-small, so a corpus indexed as text can be queried with any modality without reindexing. It was trained by freezing the v5-text-small backbone and its four task LoRA adapters and training only the cross-modal projectors over a SigLIP2 So400m vision encoder and a Whisper-large-v3 audio encoder, so text-only output is bit-identical to v5-text-small. Max sequence length 32,768; Matryoshka truncation down to 32 dimensions. Weights are CC-BY-NC-4.0.

Baidu's compact document parsing vision-language model, served by the vendor as PaddleOCR-VL-1.6-0.9B. It handles text, tables, formulas, charts, seals and text spotting through task-specific prompts, and the vendor reports a 96.33% overall score on OmniDocBench v1.6. Version 1.6 adds a region-aware data pipeline that targets weak regions found in PaddleOCR-VL-1.5 plus a progressive post-training recipe with reinforcement learning. The architecture is unchanged from 1.5, so it is a drop-in replacement.

The Allen Institute for AI's document OCR model, fine-tuned from Qwen2.5-VL-7B-Instruct on the olmOCR-mix-1025 dataset of 270,000 PDF pages and then trained with GRPO reinforcement learning against deterministic unit-test rewards. It reads a page image in a single pass and emits structured text directly: Markdown for headings and document structure, HTML for tables, and LaTeX for equations. Weights ship in BF16 and FP8 alongside the training data, toolkit and fine-tuning code.

Meta's prompt-attack classifier, an 86M-parameter model built on Microsoft's mDeBERTa-base. It labels a prompt as benign or malicious, flagging text that explicitly tries to override instructions already given to an LLM, covering both prompt injection from untrusted data and direct jailbreaks. Version 2 drops the v1 injection sub-label for a simpler binary decision, uses an energy-based loss to cut false positives on out-of-distribution text, and adds a tokenizer hardened against whitespace and fragmentation attacks.

IBM Research's safety and hallucination judge, fine-tuned from granite-4.1-8b and released in April 2026. It scores whether a prompt or response meets a criterion, emitting yes or no inside <score> tags. Pre-baked criteria cover harm, social bias, jailbreaking, violence, profanity and unethical behaviour, plus RAG context relevance, groundedness and answer relevance and function-calling hallucination. Version 4.1 adds Bring Your Own Criteria, letting callers supply arbitrary natural-language rules, and a hybrid thinking mode that trades reasoning traces for latency.

Google's image safety classifier: a 4 billion parameter model trained on Gemma 3's 4B instruction-tuned checkpoint. It takes an image plus a policy definition in the prompt and returns the probability of a 'Yes' or 'No' token, so a caller can set its own threshold. Google's card supplies three policies out of the box: sexually explicit content, dangerous content, and violence/gore. It is recommended as an input filter in front of vision-language models or as an output filter on image generation systems.

Qianfan-OCR is a 4B-parameter end-to-end document-intelligence model from Baidu's Qianfan team, published under Apache 2.0. It unifies document parsing, layout analysis and document understanding in one vision-language architecture: a Qianfan-ViT encoder (24 layers, AnyResolution up to 4K, 256 visual tokens per 448x448 tile) feeds a Qwen3-4B language model through a two-layer MLP adapter, converting images straight to Markdown. An optional Layout-as-Thought phase, triggered by think tokens, first emits bounding boxes, element types and reading order before the final output. Supported tasks span parsing, tables, formulas, charts, document QA, handwriting and key-information extraction across 192 languages.

GLM-OCR is Z.ai's compact OCR and document-parsing model, launched 3 February 2026. It pairs the in-house CogViT visual encoder with a GLM-0.5B language decoder through a lightweight cross-modal connector that downsamples visual tokens, totalling 0.9B parameters, and is trained with a multi-token-prediction loss and full-task reinforcement learning. Weights are MIT-licensed on Hugging Face and it also runs as a hosted API under the model id glm-ocr. Prompting is restricted to text, formula and table recognition plus information extraction against a caller-supplied JSON schema. The official SDK adds PP-DocLayoutV3 layout analysis for end-to-end parsing.

dots.ocr is RedNote hi lab's open-weight multilingual document parser, released 30 July 2025 under the MIT licence. A single vision-language model built on a 1.7B-parameter LLM handles layout detection and content recognition together, switching task by prompt instead of chaining separate detection and recognition stages. It returns one JSON object per page in which each element carries a bounding box, a category drawn from an eleven-class layout taxonomy, and its text - formulas as LaTeX, tables as HTML, everything else as Markdown - sorted into human reading order. Prompt modes also cover layout-detection-only, plain OCR and grounded OCR within a supplied bbox.

Granite Docling is IBM Research's 258M-parameter image-text-to-text model for document conversion, released 17 September 2025 under Apache 2.0. It builds on the Idefics3 architecture, replacing the vision encoder with siglip2-base-patch16-512 and the language model with a Granite 165M LLM, and emits DocTags that convert directly into DoclingDocuments. Its instruction set covers full-page conversion plus targeted chart-to-table, formula-to-LaTeX, code and OTSL table extraction, along with bbox-guided region inference and question answering about document structure. It is intended to be used through the Docling library.