ChartGemma

ChartGemma is a 3B chart specialist built by instruction-tuning PaliGemma (SigLIP vision encoder plus Gemma-2B) on data generated directly from chart images rather than from their underlying data tables. That choice is the point: real-world charts carry trends, annotations and visual quirks that tables do not, and the authors report state-of-the-art results across five chart benchmarks covering question answering, summarization and fact-checking.

At 3B parameters it is small enough to run on a single GPU, and the released checkpoints, dataset and Hugging Face demo make it easy to test on your own charts. Qualitative studies in the paper suggest it writes more faithful summaries of complex real-world charts than larger contemporaries. The limits are familiar for academic releases: all headline results are author-reported, adoption outside research is modest, and it will still misread unusual chart types. For chart QA and figure understanding in documents, it is a strong open baseline.

• State-of-the-art on 5 chart benchmarks (chart summarization, QA, fact-checking), author-reported (arXiv 2407.04172); no independent eval found