Google DeepMind logo

Gemma 3 27B for real-time video queries

Gemma 3 27B · Google DeepMind · Multimodal · Open-weight

Gemma 3 27B is Google DeepMind's multimodal flagship from the Gemma 3 generation, released in March 2025 with a SigLIP vision encoder and support for more than 140 languages. Gemma 3 27B is not currently in Overshoot's live model catalog, though its Gemma 4 successors are. When a model is available through the API, a live camera or screen-share stream can be queried through the same OpenAI-compatible chat-completions interface used for static images.

Gemma 3 27B is not currently in the live Overshoot model catalog (verified against the live model catalog on 2026-07-14). Availability changes over time - query GET /v1beta/models for the current list.

Developer
Google DeepMind
Parameters
27B dense
Context window
128K tokens
License
Gemma Terms of Use
Released
Mar 2025
Vision encoder
SigLIP
Overshoot availability
Not in live catalogas of 2026-07-14

What Gemma 3 27B is good at

The SigLIP vision encoder gives Gemma 3 27B strong grounding on fine visual detail, which shows up in tasks like reading small text in a frame, describing cluttered scenes, and answering questions about spatial relationships between objects. It handles single images and short sequences of frames well.

Support for 140+ languages makes it a practical choice for multilingual products that need to describe or answer questions about video for users outside English-speaking markets, without switching to a separate model per locale.

  • Fine-grained visual detail and small on-screen text
  • Multilingual visual question answering across 140+ languages
  • General scene description for camera and screen-share feeds

Gemma 3 27B and Overshoot's live video API

Gemma 3 27B is not currently in Overshoot's live model catalog. The catalog changes over time, and GET /v1beta/models is the authoritative list; its successors Gemma 4 31B and Gemma 4 26B-A4B are served on Overshoot's hosted fast path, so teams that want a Gemma model on live video today have a direct path.

On Overshoot, you publish a video track over WebRTC through LiveKit to open a Stream, then send a chat-completions request whose image_url or video_url is an ovs:// reference. The platform retains 600 seconds of frame history, so a reference can point at the latest frame, an exact timestamp, or a recent segment sampled at up to 1 fps, with answers streamed token by token over SSE and a consistent thread_id keeping prompt cache hits across repeated questions.

How it compares to Gemma 4

Gemma 3 27B predates the Gemma 4 line and sits a generation behind Gemma 4 31B and Gemma 4 26B-A4B in raw capability, but it remains a solid, well-understood choice with a mature ecosystem and the same 128K context window as its successors. Teams already deployed on Gemma 3 27B can typically move to Gemma 4 31B for a capability upgrade without changing how they structure requests.

Frequently asked questions

Can Gemma 3 27B analyze live video?

Gemma 3 27B handles single images and short frame sequences well, so the model can reason about frames sampled from live video. It is not currently in Overshoot's live model catalog; the catalog changes over time, so check GET /v1beta/models. Its Gemma 4 successors are hosted on Overshoot for live WebRTC streams.

Is Gemma 3 27B open weight?

Yes. It ships under the Gemma Terms of Use, allowing weights to be downloaded and self-hosted under Google's usage conditions.

Is Gemma 3 27B available on Overshoot?

Not currently. Gemma 3 27B is not in Overshoot's live model catalog at the moment; the catalog changes over time, and GET /v1beta/models returns the current list. The hosted lineup does include its successors Gemma 4 31B (google/gemma-4-31B-it) and Gemma 4 26B-A4B (google/gemma-4-26B-A4B-it).

What languages does Gemma 3 27B support?

Gemma 3 27B was trained with support for more than 140 languages, which makes it a reasonable default for multilingual visual question answering rather than needing a separate model per locale.

Related models