Jina VLM
Jina VLM is a 2.4-billion-parameter model for multilingual visual questions, document understanding and text generation. It joins a SigLIP2 vision encoder to a Qwen3 language decoder.
How the model works
Overlapping image tiles preserve details in high-resolution inputs. An attention-pooling connector reduces the number of visual tokens by a factor of four while retaining spatial information. This differs from CLIP-style retrieval: VLM generates an answer, while CLIP returns vectors used to find related items.Jina VLM model details
Use and access
Jina publishes local inference examples, model weights and an OpenAI-compatible beta API. The package supports multiple images as well as text-only requests. The beta endpoint is a serving status, not evidence that its weights are unavailable. Its CC BY-NC 4.0 license covers the weight download. The separate OCR model is more narrowly designed to transcribe and structure pages.Jina VLM model detailsJina VLM technical overview
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | |
| Access | API, Downloadable |
| Model identifiers | jina-vlmjinaai/jina-vlmjina-vlm-mlx |
| Parameters | 2.4 billionJina VLM model details |
| Weight license | CC BY-NC 4.0Jina VLM model details |
| Base model | Qwen3-1.7B-Base and SigLIP2-So400MJina VLM 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 - Jina CLIP v2
Jina AI · Jina CLIP
A multilingual text-and-image encoder with adjustable vector dimensions.
Launch date API · Downloadable