Doubao-Seed-1.8: BytePlus Multimodal Deep Thinking and Video Input Configuration

One-sentence takeaway

With a fixed seed-1-8-251228 version, an explicit multimodal message structure, and observable reasoning/finish fields, Seed1.8 video understanding can be turned from “chat” into a verifiable analysis workflow.

Use cases

• Suitable tasks: mixed image/video/text input, long-video summarization, visual evidence analysis, deep thinking, and structured result returns.

• Unsuitable tasks: setting max_tokens and max_completion_tokens at the same time, or treating the reasoning_content in the page examples as a field that can be unconditionally exposed to end users.

• Supported model version: BytePlus explicitly supports seed-1-8-251228.

• Supported client, Agent, or API: the BytePlus LAS/ModelArk Chat or Responses interface.

• Recommended reasoning setting and parameters: deep thinking is enabled by default; use thinking to explicitly control enabled/disabled/auto. The page recommends adjusting only one of temperature or top_p; the example ranges are temperature 0–2 and top_p 0–1.

Ready-to-use content

The following is a minimal video request template organized from the official fields. The example URL is only a placeholder to replace with an accessible video of your own; it is not a verbatim copy of the page's code.

curl --location "${LAS_ENDPOINT}/api/v1/chat/completions" \
  --header "Content-Type: application/json" \
  --header "Authorization: Bearer ${LAS_API_KEY}" \
  --data '{
    "model": "seed-1-8-251228",
    "messages": [{
      "role": "user",
      "content": [
        {"type": "video_url", "video_url": {"url": "https://<your-video-url>"}},
        {"type": "text", "text": "List the key events with timestamps, then separate observations from inferences."}
      ]
    }],
    "max_completion_tokens": 10000,
    "thinking": {"type": "enabled"},
    "response_format": {"type": "text"}
  }'

Downstream systems should save the following fields: choices[].message.content, reasoning_content (if the service allows it), finish_reason, and the response model and usage. For image input, replace video_url with the corresponding image content type; do not mix string and array formats in the same field.

Testing/workflow steps

1. Fix seed-1-8-251228, the video URL, the prompt, and max_completion_tokens, then run once with thinking enabled.

2. Check whether the output includes content, reasoning_content, and usage; after enabling the environment switch, also record finish_reason and references.

3. Split the same video into 1–5 FPS segments or submit requests in time windows, then compare tokens, latency, and event recall.

4. Repeat the test with thinking disabled/auto; change only the thinking switch, without adjusting temperature and top_p at the same time.

5. If tools are integrated, use tools and tool_choice=auto/required, validate the tool parameters before execution, and carry the tool return value into the next message.

Raw evidence and data

• BytePlus documentation lists three input types—image, video, and text—and accepts URLs, base64, and binary; the output includes the final answer llm_result and reasoning_content by default.

• thinking_type can control enabled/disabled/auto; the Chat page says deep thinking is enabled by default and supports the thinking request field.

• The supported model version is written as seed-1-8-251228; max_completion_tokens ranges from 0–64k and cannot be set at the same time as max_tokens.

• The page lists the tools, tool_choice, and response_format fields; tool_choice can be none, required, or auto, or can specify a tool.

• The official video example uses “What is the video about?” as the input, and the response returns a text summary, reasoning content, and token usage at the same time; the example usage is prompt 3930, completion 294, and reasoning 189.

Scope boundaries

• This is an interface description for BytePlus LAS/ModelArk; the model IDs, limits, and prices for Volcano Engine Ark, BytePlus, and third-party proxies must not be mixed.

• Reasoning content may involve internal reasoning or sensitive data; production UIs should display only filtered summaries and retain necessary fields as required by the protocol.

• The page does not provide complete accuracy tests for long videos; video-tool segmentation strategy, FPS, and the maximum video token count must be retested with business data.

• The specific schema for response_format must be checked against the Chat API version; the text example must not automatically be treated as a JSON Schema example.

Source excerpts or observations (for compliant short quotations only)

The official documentation describes this operator as handling image/video/text simultaneously and using thinking_type to make trade-offs between quality and performance; this template retains its field boundaries and flags proxy differences.