Gemma4-BLIP3o-Captioner-5B

Gemma4-BLIP3o-Captioner-5B is a fine-tuned image captioning model built on top of Gemma-4-E2B-it, designed to mimic and replicate the BLIP3o (Bootstrapped Language-Image Pretraining) Captioning System through targeted fine-tuning on ~3K samples from the BLIP3o-Pretrain-Long-Caption dataset. The model features a modified chat template with a hardcoded expert system prompt engineered for dense, detail-rich image captioning — covering a wide range of image categories including scenery, natural environments, portraits, objects, and more — with thinking mode disabled by default to prioritize low-latency captioning outputs. It supports sequential video frame captioning, making it suitable for temporally ordered visual description tasks. As a captioning-specialized variant, the model may produce artifacts or degraded outputs when used for general-purpose conversational or instruction-following tasks outside its captioning scope. Ideal for applications requiring structured, verbose, and contextually accurate image descriptions in privacy-focused or local inference environments, leveraging the efficient multimodal backbone of Gemma-4-E2B-it.

Note: This model is experimental and not an all-purpose one.

Model Files

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp

Base model