Step-3 for real-time video analysis

What Step-3 is good at

Step-3’s defining feature is that its mixture-of-experts routing and attention design were built together with the inference stack that serves it, rather than adapted afterward. That co-design keeps per-token decoding cost low even though the model draws from a 321B-parameter pool, which matters for any application sending frequent vision queries.

With only 38B parameters active per token, Step-3 delivers flagship-class visual understanding, including general scene description, object identification, and everyday visual question answering, at a serving cost closer to a much smaller dense model. That efficiency is the main reason to reach for it over larger, more expensive alternatives.

• Cheap per-token decoding at flagship parameter scale
• General visual question answering and scene description
• Apache 2.0 license for unrestricted self-hosting

Step-3 and the Overshoot workflow

Step-3 is not currently in Overshoot's live model catalog; GET /v1beta/models returns the current list. The workflow it would slot into is the standard one: publish a camera or screen share over WebRTC to open a Stream, then send a chat-completions request whose image_url or video_url is an ovs:// reference to the latest frame, a specific timestamp, or a recent segment sampled at 1 fps. Overshoot retains 600 seconds of frame history so recent moments stay queryable without your app buffering video.

Because Step-3 was built for cheap decoding, it is the kind of model that pairs well with high-frequency polling patterns, such as an agent that checks a stream every few seconds. Live models on Overshoot stream responses back over SSE, and a consistent thread_id keeps prompt-cache hits high across those repeated requests.

Step-3 versus other efficient MoE vision models

Step-3’s 64K context window is more modest than ERNIE 4.5 VL’s 128K or MiniMax-VL-01’s 4M, which reflects its focus on cheap, high-throughput decoding rather than very long context. Teams that need to hold long histories or many frames in a single request should look at MiniMax-VL-01; teams optimizing for frequent, low-cost queries over live video are a better fit for Step-3.

Frequently asked questions

Can Step-3 analyze live video?

Step-3 handles video frames and everyday visual question answering, but it is not currently in Overshoot's live model catalog. Live models answer questions about a WebRTC stream through ovs:// URLs in a chat-completions request, with responses streamed back over SSE. The catalog changes over time, so check GET /v1beta/models.

Is Step-3 open source?

Yes. Step-3 is released under the Apache 2.0 license, so its weights are freely downloadable and usable commercially, whether self-hosted or served by a provider.

Is Step-3 available on Overshoot?

Not currently. Step-3 is not in Overshoot's live model catalog today. The catalog changes over time, so check GET /v1beta/models for the up-to-date list of hosted and passthrough models.

What is Step-3’s context window?

Step-3 supports a 64K-token context window, enough for a focused conversation about a live stream with a handful of sampled frames, though shorter than some siblings built for very long context.

Related models

• HunyuanVL
• ERNIE 4.5 VL
• MiniMax-VL-01
• Qwen3-VL 235B-A22B