Query live video with ERNIE 4.5 VL

What ERNIE 4.5 VL is good at

Baidu trained ERNIE 4.5 VL jointly across text and vision from the start, rather than bolting a vision encoder onto a finished language model, which shows up as more consistent reasoning when a question mixes a chart, a caption, and a follow-up instruction. The heterogeneous MoE design routes text and image tokens through experts suited to each, keeping quality high without activating the full 424B parameters.

The optional thinking mode lets the model spend extra steps on harder problems, such as multi-panel charts, dense tables, or documents with nested structure, before committing to an answer. For live video that translates into more reliable readings of dashboards, whiteboards, and printed material captured by a camera.

• Joint text-and-vision reasoning from a single training run
• Thinking mode for charts, tables, and dense documents
• 128K context for long sessions with many sampled frames

ERNIE 4.5 VL and Overshoot's live video API

ERNIE 4.5 VL is not currently in Overshoot's live model catalog. The catalog changes over time, and GET /v1beta/models is the authoritative list of what is available right now, so check there before building around a specific model.

For any model that is available through the API, the workflow is the same: publish a camera or screen share over WebRTC to open a Stream, then reference it from a chat-completions request with an ovs:// URL, pointing at the latest frame, an exact timestamp, or a recent segment sampled at 1 fps. Overshoot keeps 600 seconds of frame history available, answers stream back over SSE, and a consistent thread_id per stream keeps prompt-cache hits high across repeated queries. If ERNIE 4.5 VL joins the catalog, its thinking mode would slot into that same request shape for harder chart and document questions.

ERNIE 4.5 VL versus other open MoE vision models

ERNIE 4.5 VL competes most directly with other large Apache-licensed MoE releases such as Step-3 and Qwen3-VL 235B. Its edge is the joint text-vision pretraining and thinking mode for document-heavy reasoning, while Step-3 focuses more on cheap, hardware-aware decoding and Qwen3-VL brings a wider range of model sizes to choose from.

Frequently asked questions

Can ERNIE 4.5 VL analyze live video?

The model itself accepts video frames as input, so it can reason about frames sampled from a live feed. It is not currently in Overshoot's live model catalog, however; the catalog changes over time, so check GET /v1beta/models for the models that can be pointed at a live WebRTC stream today.

Is ERNIE 4.5 VL available on Overshoot?

Not currently. ERNIE 4.5 VL is not in Overshoot's live model catalog at the moment. The catalog changes over time, and GET /v1beta/models returns the authoritative, up-to-date list of hosted and passthrough models.

Does ERNIE 4.5 VL support long context?

It supports a 128K-token context window, which is enough to hold an extended conversation about a stream together with a number of sampled frames and prior tool results without truncating early history.

How does ERNIE 4.5 VL's thinking mode help with live video?

Thinking mode lets the model take extra reasoning steps before answering, which helps with harder visual tasks like reading a busy dashboard or a multi-panel chart captured from a live camera or screen share.

Related models

• HunyuanVL
• Qwen3-VL 235B-A22B
• Step-3
• GLM-4.5V