PaliGemma 2

PaliGemma 2 is a Gemma-adjacent vision-language model family for multimodal experiments.

Google · Gemma

Editorial review

Reviewed byOpenSourcesAI EditorialLast updatedAugust 2026SourcesCanonical artifact, Representative model card

Model checkpoints, context windows, provider support, local runtime compatibility, and license terms can change quickly. Verify the exact model card before production or commercial use.

Best for

Vision-language app prototypes and multimodal evaluation.

Who should use it

Vision-language app prototypes and multimodal evaluation.
Builders who want local or self-hosted testing options.

Common workflows

Vision-language workflows
vision workflows
multimodal workflows

Deployment and hardware notes

The representative 3B 224px checkpoint needs ~1.8 GB at Q4_K_M (3.0B parameters) — the lightest model on this page, and the 224px variant is the cheapest to serve.

License and usage notes

Gemma Terms of Use. Open weights where released. Verify the exact model card and license terms for the checkpoint or hosted provider you use.

Strengths

Open weights where released model option for Gemma workflows.
Vision-language app prototypes and multimodal evaluation.
Pick the resolution variant to match the task: 224px for natural images, 448px or 896px when the answer depends on small text such as documents and charts.

Limitations

The pretrained checkpoints answer poorly out of the box, which is by design and the most common surprise: they expect task-specific fine-tuning. Text context is only 128 tokens on the pretrained release, so it cannot hold a conversation, and the Gemma terms carry use restrictions.
The representative 3B 224px checkpoint needs ~1.8 GB at Q4_K_M (3.0B parameters) — the lightest model on this page, and the 224px variant is the cheapest to serve.
Context window and limits: 128 tokens (text input/output — pretrained for short image-captioning tasks, not chat-length context).
Verify the exact model card, provider docs, license, and serving support before production use.

Local workflow notes

Pick the resolution variant to match the task: 224px for natural images, 448px or 896px when the answer depends on small text such as documents and charts.

Local runtimes: Ollama where supported, LM Studio where supported, llama.cpp where supported, Transformers

Platforms: Windows, macOS, Linux

Vision spec

This page covers a family of checkpoints. The figures below describe PaliGemma 2 3B (224px), its representative release — other sizes in the family differ.

MemoryVaries by size; the 3B 224px checkpoint needs ~1.8 GB at Q4_K_M (3B parameters)Image inputFixed 224x224 pixels, 256 image tokens (448px and 896px variants)Context128 tokens (text input/output — pretrained for short image-captioning tasks, not chat-length context)

Not a chat model: PaliGemma 2 is a base vision-language model built to be fine-tuned for one task — captioning, OCR, detection, segmentation — pairing a SigLIP encoder with a Gemma 2 text tower.

Sources to verify

Canonical artifact Representative model card

Related resources

Continue with model source notes, local tools, and implementation guides related to this model.

Gemma family hub AI tools directory Model selection guide Model source notes Models directory

HardwareVaries by size; the 3B 224px checkpoint needs ~1.8 GB at Q4_K_M (3B parameters)RuntimeOllama or LM Studio where supported, llama.cpp, Transformers, vLLMContext128 tokens (text input/output — pretrained for short image-captioning tasks, not chat-length context)Last updated2026

Canonical artifact →