Jina AI has open-sourced a small vision-language model that answers questions about images in multiple languages, and says it beats every other open model near its size.
The model, called jina-vlm, has 2.4 billion parameters. It pairs a SigLIP2 vision encoder with a Qwen3 language decoder, and handles images of any resolution by tiling them and pooling the resulting tokens with attention, which keeps token counts manageable instead of exploding with image size. According to the paper, jina-vlm posts state-of-the-art results on multilingual visual question answering among open vision-language models around the 2 billion parameter mark. The team also ran a leave-one-out ablation study, systematically removing categories of training data by task, domain, modality, and language to see which ones actually mattered and whether skills learned in one domain transferred to another. Model weights and code are public on Hugging Face.
That ablation study is the more useful part of the release. Most model papers ship a leaderboard number and little else; this one tries to show which training data is load-bearing and which is padding, information that is genuinely useful to anyone deciding what to feed their own fine-tune. For teams building multilingual document or image search tools who cannot run a frontier-scale model, a token-efficient 2.4B option with published training diagnostics is a real data point, not just marketing.
Worth remembering that "state-of-the-art" here means state-of-the-art among similarly small open models, not a claim about matching closed frontier multimodal systems, and the benchmarks are the authors' own choice of test set.