Google drops PaliGemma 2 vision models up to 28B parameters with 896px resolution

Google just significantly expanded its open vision-language model lineup with PaliGemma 2, a family of models that now scales up to a beefy 28 billion parameters. That’s a massive jump from the original PaliGemma, which topped out at just 3B. The new release, built by pairing the SigLIP image encoder with the latest Gemma 2 language models, comes in three distinct sizes—3B, 10B, and 28B—each capable of handling three different input resolutions: 224x224, 448x448, and 896x896 pixels. That’s nine pre-trained checkpoints to choose from, giving developers a real dial to turn when balancing inference speed against visual detail.

This isn’t just a scale-up for scale’s sake. The pre-training data recipe is genuinely diverse, pulling from massive sources like the multilingual WebLI dataset, the translated CC3M-35L collection spanning 35 languages, and object-aware question-answer pairs from OpenImages. The team’s internal benchmarks show the models can then be fine-tuned effectively for specific tasks with relatively few examples. To prove it, Google released two models fine-tuned on the DOCCI dataset, which excel at generating long, nuanced image captions that capture spatial relationships and even world knowledge. One of those DOCCI-tuned models clocks its average caption at 172 sentences, with a factual inaccuracy rate low enough to raise eyebrows.

Hugging Face has already thrown its weight behind the release, shipping full transformers integration, fine-tuning scripts, and even a demo they built themselves. Their team fine-tuned the 3B model on a slice of the VQAv2 dataset using LoRA and PEFT, a technique that keeps the memory footprint manageable for smaller labs and solo developers. That’s the real story here: a 28B vision model you can run, remix, and even commercialize under the permissive Gemma license. The days of closed, API-gated multimodal models are starting to look numbered.

Comparing the DOCCI-tuned variants to other captioning models, the PaliGemma 2 checkpoints show a clear tradeoff between verbosity and factual precision, measured by a metric called Non-Entailment Sentences. It’s a more honest approach than just reporting BLEU scores. The question now is whether the community will push these models beyond captioning into agentic use cases—can a model this large reason about a 896x896 screenshot well enough to navigate a UI? The weights are open, so we’re about to find out.