Doubao Big Model 1.6 - A multimodal deep thinking model launched by ByteDance.

What is the Doubao large model 1.6?

Doubao-Seed-1.6 is a multimodal deep thinking model launched by ByteDance. The model supports three thinking modes: auto, thinking, and non-thinking. It can handle various input formats, including text, images, and videos, and supports high-quality text output. The model features a 256k-long context window, a maximum input length of 224k tokens, and an output length of up to 16k tokens, demonstrating powerful reasoning capabilities. The model has performed excellently in multiple authoritative evaluations, showing significant improvements in reasoning and mathematical abilities. Doubao 1.6 is widely used in content creation, intelligent dialogue, code generation, and other fields, providing enterprises and developers with a powerful AI productivity tool. Doubao 1.6 is available on Volcano Engine, and enterprises and developers can experience it via API.

The latest upgrade to Doubao's large model 1.6 innovatively introduces four thinking length modes: Minimal, Low, Medium, and High. This precisely balances performance, latency, and cost, making it the first large model in China to natively support 'tiered adjustment of thinking length'. Among them, the low thinking length mode reduces total output tokens by 77.5% and thinking time by 84.6% compared to the previous version, while maintaining the same performance, significantly improving efficiency.

Main functions of the large bean bun model 1.6

• reasoning abilitySignificantly improved inference speed, accuracy, and stability, enabling it to support the implementation of more complex business scenarios.
• Think and search simultaneously with DeepResearchIt possesses the ability to search while thinking, and can provide recommendations based on missing information after multiple rounds of thinking and searching. The DeepResearch feature supports the rapid generation of research reports.
• Multimodal understanding abilityThe entire series natively supports multimodal thinking capabilities, enabling the understanding and processing of various modal data such as text, images, and videos.
• Graphical interface operation capability (GUI operation)Based on visual depth thinking and precise positioning, it can interact and operate with browsers and other software to efficiently execute various tasks.

Three models of Doubao large model 1.6

• doubao-seed-1.6This is an all-around comprehensive model, and the first thinking model in China to support 256K contexts. The model possesses multiple capabilities, including deep thinking, multimodal understanding, and graphical interface operation. Users can flexibly choose to enable or disable the deep thinking function according to their needs, and it supports an adaptive thinking mode. The adaptive mode automatically determines whether to enable deep thinking based on the difficulty of the prompts, improving effectiveness while significantly reducing token consumption.
• doubao-seed-1.6-thinkingDeep Thinking Enhanced Edition is an enhanced version of the Doubao Big Model 1.6 series, focusing on deep thinking. The model further improves its fundamental capabilities in coding, mathematical calculation, and logical reasoning, enabling it to handle more complex tasks. It also supports 256K of context, allowing it to understand and generate longer text content, making it suitable for scenarios requiring deep analysis and complex reasoning.
• doubao-seed-1.6-flashThis is the ultra-fast response version, the fastest version in the Doubao Big Model 1.6 series. It features deep thinking and multimodal understanding capabilities and supports 256K context. The model has extremely low latency, with a Top-of-Pipeline Time (TOPT) of only 10ms, enabling rapid response to user requests. The model's visual understanding capabilities are comparable to flagship models from competing companies, making it suitable for scenarios with extremely high responsiveness requirements, such as real-time interaction and visual task processing.

Performance of Doubao Large Model 1.6

• GPQA Diamond TestThe Doubao 1.6-thinking model achieved a score of 81.5, reaching the top tier globally and is one of the best reasoning models currently available.
• AIME 25 Mathematics AssessmentThe Doubao 1.6-thinking model scored 86.3 points, an improvement of 12.3 points compared to the Doubao 1.5 deep thinking model.

The pricing model of Doubao's large model 1.6

Doubao's big model 1.6 is based on a unified pricing model. Regardless of whether the deep thinking mode is enabled or not, and regardless of whether the input is text or visual, the price of tokens is the same.

• Input length 0-32K:
  • Enter price0.8 yuan/million tokens.
  • Output Price8 yuan per million tokens.
• Input length 32K-128K:
  • Enter price1.2 yuan/million tokens.
  • Output Price16 yuan per million tokens.
• Input length 128K-256K:
  • Enter price2.4 yuan per million tokens.
  • Output Price24 yuan per million tokens.
• Input 32K, output up to 200 tokens:
  • Enter price0.8 yuan/million tokens.
  • Output Price2 yuan per million tokens.

How to use the Doubao large model 1.6

• Register and log in to the Volcano Engine platformVisit the official website of Volcano Engine and follow the prompts to register and log in.
• Launching Doubao Large Model Service: Go to the service page and find the service page for Doubao Big Model 1.6.
• Activate service: Activate the Doubao Big Model 1.6 service according to the page prompts.
• Select versionChoose the appropriate model version according to your needs (such as doubao-seed-1.6, doubao-seed-1.6-thinking, or doubao-seed-1.6-flash).
• Get API KeyAfter the service is activated, the platform provides an API key, which is used for authentication when calling the model.
• Calling Model:
  • Using API interfaceThe Doubao Big Model 1.6 is based on API calls.
  • Build Request: Build request data based on requirements, including input text, parameter settings, etc.
  • Send RequestUse HTTP requests to send data to the model's API interface.
  • Receive responseAfter the model finishes processing, it returns response data, including the generated text or other results.
• Example code (Python)The following is sample code for using Python to call the Doubao Big Model 1.6:

import requests
import json

# API密钥和接口地址
api_key =
"your_api_key"
api_secret =
"your_api_secret"
model_version =
"doubao-seed-1.6"
# 或doubao-seed-1.6-thinking、doubao-seed-1.6-flash
api_url =
f"https://api.volcengine.com/v1/model/{model_version}"
# 请求数据
data =
{
"input":
"你的输入文本",
"parameters":
{
"max_length":
256,
# 输出的最大长度
"temperature":
0.7,
# 随机性参数
"top_p":
0.9,
# 核心采样参数
"top_k":
50,
# 核心采样参数
"do_sample":
True
# 是否采样
}
}
# 设置请求头
headers =
{
"Authorization":
f"Bearer {api_key}",
"Content-Type":
"application/json"
}
# 发送请求
response = requests.post(api_url, headers=headers, data=json.dumps(data))
# 处理响应
if response.status_code ==
200:
 result = response.json()
print("模型输出：", result["output"])
else:
print("请求失败，状态码：", response.status_code)
print("错误信息：", response.text)

Project address for Doubao Large Model 1.6

• Project official website:https://console.volcengine.com/ark/region:ark+cn-beijing/model/detail?Id=doubao-seed-1-6

Application scenarios of the large bean bun model 1.6

• Content creationGenerates advertising copy, news reports, stories, novels, etc., helping users quickly produce high-quality content.
• Intelligent dialogueUsed in intelligent customer service and chatbots, it provides a natural and smooth multi-turn dialogue experience, improving user interaction efficiency.
• Code generationGenerates front-end code snippets based on requirements, assisting developers in troubleshooting errors and improving development efficiency.
• Educational guidanceIt answers subject-specific questions, generates teaching resources, and assists students in learning and teachers in lesson preparation.
• Multimodal content generationCombine image or video input to generate relevant text descriptions or creative content, facilitating multimedia creation.