PaliGemma 2 28B on live video streams

What PaliGemma 2 28B is good at

PaliGemma 2 28B is the go-to open base when a team needs to fine-tune a vision model for a narrow, well-defined task rather than deploy a general chat assistant. Its prefix-LM design conditions generation on a task prompt, which makes it a natural fit for captioning, object detection, and segmentation transfer once adapted to a target domain.

The tradeoff for that specialization is a short 512-token text context, so it is not built for long multi-turn conversation. Instead it excels at short, structured outputs, bounding boxes, region descriptions, or segmentation masks, produced consistently frame after frame.

• Image and video-frame captioning at scale
• Object detection and localization transfer after fine-tuning
• Segmentation mask generation for downstream pipelines

PaliGemma 2 28B and Overshoot's streaming API

PaliGemma 2 28B is not currently in Overshoot's live model catalog; GET /v1beta/models returns the current list. The streaming workflow it would fit is the same for every model on the platform: publish a camera or screen-share track over WebRTC through LiveKit to open a Stream, then reference frames in a chat-completions request with an ovs:// URL: the latest frame, a specific timestamp, or a recent segment. Overshoot keeps 600 seconds of frame history available for these references.

A fine-tuned specialist like PaliGemma 2 28B suits pipelines that call a model repeatedly against a live feed for detection or captioning at a steady cadence, so a deployment of this kind pairs naturally with SSE streaming and low time to first token wherever it is served.

How it fits alongside Gemma

PaliGemma 2 28B occupies a different niche from the general chat-oriented Gemma models: where Gemma 3 27B and Gemma 4 31B are built for open-ended conversation and reasoning about a frame, PaliGemma 2 28B is a specialist base meant to be fine-tuned onto one task at a time. Teams often prototype with a chat-capable Gemma model and move to PaliGemma 2 28B once a specific detection, captioning, or segmentation task is well defined and worth fine-tuning for.

Frequently asked questions

Can PaliGemma 2 28B analyze live video?

PaliGemma 2 28B works on individual frames, so it can analyze video by processing sampled frames when self-hosted. It is not currently in Overshoot's live model catalog, so it cannot be referenced against an Overshoot WebRTC stream today. The catalog changes over time; check GET /v1beta/models for the current list.

Is PaliGemma 2 28B open weight?

Yes. It ships under the Gemma Terms of Use, which permits downloading and self-hosting the weights, including for fine-tuning onto a custom task.

What is PaliGemma 2 28B best used for?

It is a prefix-LM vision model built for fine-tuning, not general chat, and is the go-to open base for captioning, object detection, and segmentation transfer once adapted to a specific task.

Why does PaliGemma 2 28B have such a short context window?

Its 512-token text context reflects its design as a task-specific specialist rather than a conversational model. Outputs are typically short and structured, such as a caption, bounding box, or segmentation mask, so a long context window is not the priority.

Related models

• Florence-2
• Gemma 3 27B
• Gemma 4 31B
• Qwen2.5-VL 7B