Gemma 4 26B-A4B: sparse vision for live streams

What Gemma 4 26B-A4B is good at

Because only 4B of its 26B parameters activate per token, Gemma 4 26B-A4B serves at a fraction of the cost of a dense model of comparable quality. That makes it well suited to applications that query video continuously rather than occasionally, such as monitoring dashboards or assistants that re-check a scene every few seconds.

Overshoot published a public engineering study on batching this model’s vision encoder that kept p95 time to first token near 120ms at 20 requests per second, using six 480p frames per request. That result reflects why 26B-A4B, not the dense 31B model, is the default choice when a product needs to sustain high query volume against live streams.

• High-frequency scene monitoring and change detection
• Cost-sensitive live-video assistants at meaningful request volume
• Batched vision-encoder workloads where throughput matters as much as latency

Running Gemma 4 26B-A4B on Overshoot

Publish video over WebRTC via LiveKit to open a Stream, then send chat-completions requests that reference frames with an ovs:// URL: the latest frame, an exact timestamp, or a recent segment sampled at up to 1 fps. Point the model id at google/gemma-4-26B-A4B-it to reach this specific variant.

Responses stream over SSE with the low time to first token that the batching study measured. Reuse a thread_id across repeated queries against the same stream to keep hitting the prompt cache instead of reprocessing shared context on every call.

How it compares to Gemma 4 31B

Gemma 4 26B-A4B and Gemma 4 31B are the two faces of the same generation, released together in April 2026 with the same 128K context window. 31B is dense and simpler to reason about at a fixed per-token cost, while 26B-A4B’s sparse routing keeps active compute low, which is why it is the model Overshoot optimized for sustained, high-throughput live-video traffic.

Frequently asked questions

Can Gemma 4 26B-A4B analyze live video?

Yes. Overshoot streams live video over WebRTC and lets a chat-completions request reference the latest frame or a recent segment with an ovs:// URL. Gemma 4 26B-A4B, served as google/gemma-4-26B-A4B-it, reads the referenced frames and streams its answer back over SSE.

Is Gemma 4 26B-A4B open weight?

Yes. It ships under the Gemma Terms of Use, which allows downloading and self-hosting the weights subject to Google’s usage conditions. That open-weight status is why Overshoot can run it on its low-latency hosted fast path.

How fast is Gemma 4 26B-A4B on Overshoot?

Overshoot’s own batching study measured p95 time to first token near 120ms at 20 requests per second using six 480p frames per request. In general, Overshoot-hosted open-weight models answer in about 200ms, and 26B-A4B’s 4B active parameters keep it toward the faster end of that range.

What is the difference between Gemma 4 26B-A4B and Gemma 4 31B?

Gemma 4 26B-A4B is a mixture-of-experts model with 26B total and 4B active parameters per token, while Gemma 4 31B is a dense 31B model. Both share a 128K context window and April 2026 release date, but 26B-A4B is the variant tuned for low-cost, high-volume serving.

Related models

• Gemma 4 31B
• Gemma 3 27B
• Qwen3-VL 32B
• Phi-4 Multimodal
• Qwen3.6 35B A3B
• Qwen3.6 27B