SeaWhale AI

English

简体中文

English

简体中文

Appearance

Sign in

Doubao-Seed-2.1-turbo

doubao-seed-2-1-turbo-260628
Get API keyAPI reference

The Seed2.1 series is a low-cost, low-latency version designed for large-scale production scenarios, with a full feature set and performance comparable to Doubao-Seed-2.1-pro. It fully supports Coding, Agent, GUI tasks, and image and video understanding, and offers a 256K context window and a maximum output of 256K, making it suitable for reliably handling a large volume of online calls.

Context window256K
Provider豆包
Released2026/10/20

Playground

Pricing

The same model is available through multiple service channels — choose based on latency, reliability and cost.

Prices in $ / 1M tokens
To pick a channel, add a provider field to the request body, for example "provider": { "channel": "direct" }. Valid values are direct / stable / economical; omit it to use the default channel.

Direct

Direct upstream connection — best when you need native behavior and the full context window.

InputOutputCache readCache write
0.45/M2.24/M0.09/M<0.01/M

Overview

Input
Text Image
Output
Text

Doubao-Seed-2.1-turbo API: A low-cost, low-latency version with performance comparable to Pro, suitable for scaled production calls

Doubao-Seed-2.1-turbo is the production-at-scale tier in ByteDance's Doubao Seed2.1 series. It was released alongside the flagship Doubao-Seed-2.1-pro at the Volcano Engine FORCE conference on June 23, 2026, with version number 260628. Volcano Ark's official positioning for it is "a low-cost, low-latency version for scaled production scenarios, fully featured, with performance comparable to doubao-seed-2-1-pro," designed to reliably handle large volumes of online calls for Coding, Agent, and multimodal understanding tasks.

According to benchmarks published by Seed official, the gap between Turbo and Pro is concentrated in long-chain Agent and video action perception; most other benchmarks differ by only 1–3 points, and some even outperform: math reasoning BeyondAIME 88.0 (Pro is 87.0), office-related Workspace Bench 54.7 (Pro is 53.0). It retains almost all capabilities, while the official unit price is only half that of Pro, making it the most cost-effective general-purpose tier in the Seed2.1 series.

Two things to know in advance: the context window is 256K (the Pro 260915 version is 1024K), so be careful with ultra-long document scenarios; although the maximum response is 256K, the default output is only 4K, so long-output tasks must explicitly set max_tokens. Deep thinking is enabled by default, and the chain of thought is counted in the output; when thinking is enabled, leave enough headroom for max_tokens.

SeaWhale AI provides Doubao-Seed-2.1-turbo through an OpenAI-compatible interface, supporting image and video input, deep thinking, tool calling, and structured output.

Get API key · Model ID: doubao-seed-2-1-turbo-260628

Why choose Doubao-Seed-2.1-turbo

  • Performance comparable to Pro — Terminal Bench 2.1 67.6 (Pro 71.0), MMMU-Pro 80.1 (Pro 81.6); most benchmark gaps are within 3 points
  • Math reasoning outperforms Pro — BeyondAIME 88.0, higher than Pro's 87.0 and Claude Opus 4.7's 79.0
  • Strong spatial and visual perception — ERQA 71.3 and BabyVision 62.9, both higher than GPT-5.5 (64.5, 55.9) in the official comparison
  • Long-video understanding — VideoMME 89.0, higher than Gemini 3.1 Pro's 86.7 in the official comparison
  • 256K context + 256K output — maximum response and maximum chain of thought are both 256K
  • Half the unit price of Pro — significantly lower cost for large-batch online calls and multi-turn Agent calls

Core capabilities

01 Production-scale Coding and Agent

Turbo retains the full Coding and Agent capabilities of Seed2.1: full-cycle delivery from requirements analysis, feature implementation, issue fixing, environment setup, and validation, as well as GUI task handling and tool calling. In official benchmarks, Terminal Bench 2.1 scores 67.6, NL2Repo-Bench 43.7, and Workspace Bench 54.7, making it suitable as the agent execution layer in high-concurrency scenarios.

  • Multi-step engineering tasks in terminal environments
  • Batch code modification, code review, and test generation
  • GUI automation and tool-calling workflows

02 Multimodal understanding

Turbo's multimodal scores are very close to Pro: MathVision 90.1, MMMU-Pro 80.1, CharXiv-RQ 82.5, ERQA 71.3, and MMLongBench-128K 76.9. Visual knowledge WorldVQA 48.6 is higher than GPT-5.5 (34.6) and Claude Opus 4.7 (35.9).

  • Understanding and extraction of charts, tables, and document images
  • Spatial reasoning and interface screenshot understanding
  • Multi-image long-context Q&A

03 Video understanding

For video, official benchmarks are VideoMME 89.0, OVOBench (streaming video) 79.2, VideoSimpleQA 71.4, and Minerva 65.9; long-video and streaming-video understanding are both in the top tier, suitable for large-scale video content moderation, retrieval, and summarization.

  • Long-video summarization and clip retrieval
  • Real-time Q&A on streaming video
  • Video content moderation and tag generation

04 Reasoning and engineering features

Deep thinking is enabled by default, and the level of thinking can be adjusted via reasoning_effort. BeyondAIME 88.0 shows that it is not weaker than Pro in mathematical and logical reasoning. It also supports function calling, structured output (the json_schema mode is recommended), and explicit caching such as implicit caching, prefix caching, and Session caching.

  • Adjustable thinking intensity to balance performance and latency
  • Tool calling and JSON structured output
  • Context caching reduces the cost of repeated calls with long prefixes

Best use cases

Scenario Description
High-concurrency online services Customer service, assistants, Q&A, and other high-volume calls, with cost and throughput as priorities
Batch processing Document extraction, structured output, batch code modification
Agent execution layer Pair with Pro for planning; Turbo executes large numbers of subtasks
Video content processing Long-video summarization, moderation, retrieval
Multimodal understanding Q&A and extraction for charts, screenshots, and document images
Math and reasoning tasks BeyondAIME 88.0; reasoning tasks do not need Pro

Differences between Doubao-Seed-2.1-turbo, Doubao-Seed-2.1-pro, and Doubao-Seed-2.1-lite

Capability Doubao-Seed-2.1-turbo Doubao-Seed-2.1-pro Doubao-Seed-2.1-lite
Model ID doubao-seed-2-1-turbo-260628 doubao-seed-2-1-pro-260915 doubao-seed-2-1-lite-260915
Release date June 23, 2026 September 16, 2026 (series debut June 23) September 2026
Context window 256K token 1024K token 1024K token
Max response / chain of thought 256K / 256K 256K / 256K 256K / 256K
Input modalities Text, image, video Text, image, video Text, image, video, audio
Terminal Bench 2.1 67.6 71.0 (June debut version) —
Agent Startup Bench 54.0 68.8 (June debut version) —
VideoMME 89.0 89.2 (June debut version) —
Positioning Low cost, low latency, performance comparable to Pro Series flagship, production-grade Agent and Coding Lightweight high-frequency, office and multimodal

Specifications come from the Volcano Ark model list and model release announcements; benchmark scores come from the official Seed Seed2.1 page. Pro's scores correspond to the June debut 260628 version; official independent benchmarks for Lite have not been published. Actual billing is subject to the real-time price card at the top of the page.

FAQ

When was Doubao-Seed-2.1-turbo released? Released on June 23, 2026 with the Seed2.1 series at the Volcano Engine FORCE conference, version 260628; the API went live on Volcano Ark at the same time.

How much does it differ from Doubao-Seed-2.1-pro? In official benchmarks, most gaps are 1–3 points: Terminal Bench 2.1 67.6 vs. 71.0, MMMU-Pro 80.1 vs. 81.6, VideoMME 89.0 vs. 89.2; BeyondAIME (88.0 vs. 87.0) and Workspace Bench (54.7 vs. 53.0) are slightly higher instead. The clear gaps are in long-chain Agent (Agent Startup Bench 54.0 vs. 68.8) and video action perception (TOMATO 56.8 vs. 79.5).

What are the context and output limits? Context and maximum input are both 256K tokens, maximum response is 256K tokens, and maximum chain of thought is 256K tokens. The default response length is only 4K; explicitly set max_tokens when long output is needed.

How do I configure deep thinking? Deep thinking is enabled by default; adjust the level of thinking via reasoning_effort. Keep a higher setting for complex coding and reasoning tasks; for latency-sensitive tasks such as dialogue, classification, and extraction, lower it to reduce output tokens and waiting time.

What tools and features are supported? Supports image and video input, function calling, structured output (the json_schema mode is recommended), GUI task handling, and explicit caching such as implicit caching, prefix caching, and Session caching. Audio input is not supported; if you need to directly understand speech, choose Doubao-Seed-2.1-lite.

Where does Doubao-Seed-2.1-turbo fall behind? Yes. First, long-chain Agent is clearly weaker than Pro: Agent Startup Bench 54.0 is 14.8 points lower than Pro, and xDailyBench 56.4 is also lower than Pro's 61.0. Second, code debugging and long-horizon repository development are weaker: SWE-Atlas Debugging 30.6 and NL2Repo-Bench 43.7, lower than GPT-5.5 (44.7) and Claude Opus 4.7 (58.2). Third, video action perception TOMATO 56.8 is far lower than Pro's 79.5. Fourth, the context is only 256K, so ultra-long documents need to be split or switched to Pro. On knowledge, SuperGPQA 67.4 is also lower than Gemini 3.1 Pro's 76.6.

Should I choose Turbo, Pro, or Lite? Choose Turbo for high-volume online calls, when cost and latency are priorities, and when 256K context is sufficient; choose Pro when you need 1024K ultra-long context or the most complex long-chain Agent tasks; choose Lite for high-frequency lightweight office work or when audio input is needed. All three can be compared directly under the same SeaWhale AI account using the same code.

Why choose SeaWhale AI to use the Doubao-Seed-2.1-turbo API

  • No Volcano Engine account required — call directly with one key; no separate application for Ark model access
  • OpenAI-compatible interface — multimodal request format is consistent with OpenAI; just change the model name in existing code
  • Compare within the same series under one account — share a key with Doubao-Seed-2.1-pro, 2.1-lite, and other domestic flagship models
  • Pay-as-you-go — billed by actual token usage, no minimum spend

API

API integration ​

Model IDUse this value as the model in inference requests
doubao-seed-2-1-turbo-260628
API KeyBearer token used to authenticate inference requests
Manage API keys
Base URLOpenAI compatible · /chat/completions
OpenAIhttps://api.haijingai.com/v2/
provider OptionalSelects a service channel; omit it and the system picks the default
"provider": { "channel": "direct" }

doubao-seed-2-1-turbo-260628 usage examples ​

SeaWhale AI is compatible with the OpenAI API protocol, so you can call it with the OpenAI SDK or plain HTTP requests. Streaming is enabled by default.

js
curl https://api.haijingai.com/v2/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <API_KEY>" \
  -d '{
    "model": "doubao-seed-2-1-turbo-260628",
    "messages": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "Hello!"}
    ],
    "provider": { "channel": "direct" },
    "stream": true
  }'
# provider is optional — remove this line to use the default channel
js
from openai import OpenAI

client = OpenAI(
    base_url="https://api.haijingai.com/v2",
    api_key="<API_KEY>",
)

stream = client.chat.completions.create(
    model="doubao-seed-2-1-turbo-260628",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Hello!"},
    ],
    stream=True,
    # Optional: pick a service channel; omit to use the default
    extra_body={"provider": {"channel": "direct"}},
)

for chunk in stream:
    if chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="", flush=True)
js
import OpenAI from 'openai'

const client = new OpenAI({
  baseURL: 'https://api.haijingai.com/v2',
  apiKey: '<API_KEY>',
})

const stream = await client.chat.completions.create({
  model: 'doubao-seed-2-1-turbo-260628',
  messages: [
    { role: 'system', content: 'You are a helpful assistant.' },
    { role: 'user', content: 'Hello!' },
  ],
  stream: true,
  // Optional: pick a service channel; omit to use the default
  // @ts-expect-error provider is a SeaWhale AI extension, not in the OpenAI SDK types
  provider: { channel: 'direct' },
})

for await (const chunk of stream) {
  process.stdout.write(chunk.choices[0]?.delta?.content ?? '')
}
Contact support