ReaderLM v2
ReaderLM v2 is Jina's 1.54-billion-parameter model for HTML parsing and extraction. It adds direct schema-based JSON output to the Markdown conversion offered by the first ReaderLM models.
How the model works
The decoder has 28 layers, a 1,536-wide hidden state and two key-value heads. Jina documents 512K tokens for combined input and output, with support for 29 languages. Contrastive loss during training targets repetitive or unstable long outputs. The model can preserve elements such as nested lists, code blocks, tables and equations.ReaderLM v2 model details
Use and access
Weights are published under CC BY-NC 4.0. The model card also documents use inside Reader with the readerlm-v2 engine setting. That engine is one part of the Reader service: fetching a URL, rendering a page and extracting content are not all properties of the downloaded language model. Its input is HTML rather than a camera image.ReaderLM v2 model detailsReaderLM v2 technical overview
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | |
| Access | API, Downloadable |
| Model identifiers | ReaderLM-v2jinaai/ReaderLM-v2 |
| Context window | 512,768 tokens Exact max_position_embeddings value in the published configuration; the model card abbreviates this as 512K. Input and generated output share the context. ReaderLM-v2 published model configuration |
| Parameters | 1.54 billionReaderLM v2 model details |
| Weight license | CC BY-NC 4.0ReaderLM v2 model details |
Related models
- Jina OCR v1
Jina AI · Jina OCR
A document parser that converts page images into structured text.
Launch date API · Downloadable - ReaderLM 0.5B
Jina AI · ReaderLM
A small HTML-to-Markdown model for local page-content conversion.
Launch date Downloadable - ReaderLM 1.5B
Jina AI · ReaderLM
The larger first-generation ReaderLM model for HTML-to-Markdown conversion.
Launch date Downloadable