Llama 4 Maverick on real-time video

What Llama 4 Maverick is good at

Maverick was trained end to end as a natively multimodal model, so it reasons over images and text with the same weights rather than a bolted-on vision adapter. Drawing 17B active parameters from a 400B expert pool gives it more capacity per query than Scout, which shows up on harder visual reasoning, dense document layouts, and multi-step questions that combine several images.

The tradeoff is cost and latency relative to Scout: Maverick is the model to reach for when a task needs the extra reasoning depth, not the default for every request. Its 1M-token context window is smaller than Scout's 10M, but still comfortably covers long conversation history and many sampled video frames in a single thread.

• Complex visual reasoning across multiple images or frames
• Dense document and layout understanding
• Higher-stakes questions where answer quality matters more than cost

Llama 4 Maverick and Overshoot's streaming workflow

Llama 4 Maverick is not currently in Overshoot's live model catalog. Publishing to Overshoot works the same way regardless of which catalog model answers the question: create a Stream, publish a LiveKit video track over WebRTC, and Overshoot retains 600 seconds of frame history for that stream. A chat-completions request references the footage with an ovs:// URL, anchored to the latest frame, an exact timestamp, or a recent segment.

If Maverick were served through the API, responses would stream back over SSE token by token rather than waiting for a full completion. Passing a consistent thread_id keeps the prompt cache warm across repeated queries against the same stream, which would matter more for Maverick than for smaller models given its larger active parameter count.

Maverick versus Scout

Maverick and Scout activate the same 17B parameters per token, but Maverick draws from a 400B pool against Scout's 109B, trading some efficiency for higher ceiling quality. Teams building live video applications typically default to Scout for cost and latency, and route a subset of harder questions to Maverick when the extra reasoning depth is worth the price. Llama 3.2 Vision remains an option where its adapter-based architecture or older licensing terms are a better fit for a given deployment.

Frequently asked questions

Can Llama 4 Maverick analyze live video?

Maverick is natively multimodal and reasons over sampled video frames alongside text. It is not currently in Overshoot's live model catalog, so it cannot be called against Overshoot streams today. Catalog models answer questions about a live WebRTC stream via an ovs:// URL in a standard chat-completions request, with the answer streamed back over SSE.

Is Llama 4 Maverick open source?

Maverick ships under the Llama 4 Community License, with downloadable weights and self-hosting allowed under some commercial conditions.

Is Llama 4 Maverick available on Overshoot?

Not currently. Maverick is not in Overshoot's live model catalog. The catalog changes over time, so check GET /v1beta/models for the current list of models you can call through the API.

What is the difference between Llama 4 Maverick and Scout?

Both activate 17B parameters per token, but Maverick pulls from a 400B expert pool for higher quality while Scout uses a 109B pool and a much larger 10M-token context window. Maverick is the choice for harder visual reasoning, and Scout is the choice for cost-efficient, long-context work.

Related models

• Llama 4 Scout
• Llama 3.2 Vision
• Pixtral Large
• Qwen3-VL 235B-A22B