PaliGemma 2
Google · USA · 2024
The same vision encoder as the first PaliGemma, now bolted onto every size of Gemma 2 — up to 27.6 billion parameters — and pushed into tasks like music scores, molecular diagrams and chest X-ray reports.
PaliGemma 2 is the second generation of Google's open vision-language model, announced on 5 December 2024. The recipe of the first version is untouched — the SigLIP-So400m image encoder in front of a Gemma language model — but instead of a single 2B language model the encoder is now paired with the whole Gemma 2 range: the 2B, 9B and 27B models, giving vision-language models of 3.03, 9.66 and 27.65 billion parameters. Each size was trained at three resolutions, 224, 448 and 896 pixels. Google presents it as a drop-in replacement: anyone who was using the first PaliGemma can switch checkpoints without rewriting code. The point of the release, according to the technical report, was less about topping familiar leaderboards and more about mapping the interplay between model size, input resolution and task type — which combination is worth paying for. Alongside that, Google widened the range of transfer tasks well beyond the first generation and reports state-of-the-art results on several of them: recognising table structure, reading molecular structure diagrams, reading music scores, long fine-grained captioning and generating radiology reports from chest X-rays. That last one is worth reading carefully — it is a research result on a benchmark, not a clinical tool, and Google does not present it as one. On 19 February 2025 the family gained mix checkpoints: weights tuned on a mixture of tasks that answer short and long captioning, OCR, questions about an image, detection and segmentation out of the box, published for all three sizes at 224 and 448 pixels. Task-specific checkpoints, of which the first generation had fifty-four, are here down to two — both for long captioning on DOCCI. Licensing is unchanged from the first version: the Gemma Terms of Use rather than Apache 2.0, with gated repositories requiring accepted terms and manual approval. Traffic tells an awkward story about the drop-in upgrade: 64,773 downloads in the thirty days to 17 September 2026, against 619,679 for the model it was meant to replace.
▸News
▸Videos
No videos yet.
▸Reviews
No reviews yet. Be the first!