Robots Atlas>ROBOTS ATLAS
Doubao

Doubao

Doubao Seed 2.0 · Family: Doubao
ByteDance LLM family powering the Doubao assistant: multimodal MoE models (text, image, speech) with a deep thinking mode; the most popular AI chatbot in China.
✓ Active✓ Public accessLLMMultimodalReasoning modelTool-using model📁 Doubao
Release date
1 August 2023
Access:APIHostedDeployment:☁ Cloud

Overview

Doubao (Chinese 豆包, “little bean”) is a family of large language models developed by ByteDance that powers the Doubao AI assistant — the most popular AI chatbot in China. The assistant launched in August 2023 and reached about 60 million monthly active users by November 2024, growing to 330 million users by May 2026 (when paid subscription tiers were also introduced).

Doubao models are built on a mixture-of-experts (MoE) architecture with an integrated training-inference design. According to ByteDance, the Doubao-1.5-pro model — with activated parameters equal to about 1/7 of a dense model — surpassed that dense model’s quality; pretraining spanned roughly 9 trillion tokens.

The family is multimodal: it supports text, image understanding (with native dynamic resolution) and speech via an end-to-end Speech2Speech framework. The models also offer a “deep thinking” (reasoning) mode — the Doubao-1.5-pro-AS1-Preview variant surpassed reasoning models such as o1-preview on the AIME benchmark.

The latest generation is Doubao Seed 2.0 (stable release on 14 February 2026). The models are available through the Volcano Engine (Volcengine) platform and ByteDance apps on iOS, Android and the web; the international variant of the assistant is Cici/Dola. These are proprietary models.

Classification
LLMMultimodalReasoning modelTool-using model
Family: Doubao
Access & deployment
APIHosted
Cloud
Weights: Closed
Key parameters
Tools · ✓ Fine-tuning
📥 Input: text, image, audio

Technical specification

License
Proprietary
Hardware requirements
A hosted (cloud) model, delivered as a service via the Volcano Engine platform and ByteDance apps; MoE architecture with an integrated training-inference design. No local deployment (closed weights).
Features:Tool useFine-tuning
Modalities
⬇ Input
textimageaudio
⬆ Output
textaudiocode

Capabilities and applications

Native model capabilities
Reasoning
The model's ability to reason logically and solve complex problems.
Category: reasoning
Extended thinking mode
A reasoning-model variant with a larger inference budget: more thinking cycles, higher answer precision at the cost of response time. Choice between 'standard' and 'extended' thinking is left to the user (e.g. the selector in GPT-5.2 Pro).
Category: reasoning
Multi-step reasoning
Carrying out multi-step chains of reasoning across long, complex tasks.
Category: reasoning
Coding
Generating, analysing and modifying code in many programming languages. Covers writing functions, debugging, refactoring, code review, and creating tests. Measured by benchmarks such as HumanEval and SWE-bench.
Category: coding
Tool use
The model's ability to call external functions, APIs and tools during a conversation: calculator, search engine, code editor, database. The model decides when and how to use a tool and interprets its result.
Category: planning
Function Calling
Category: planning
Long context
Support for large context windows — tens to hundreds of thousands (or millions) of input tokens. Enables analysis of entire codebases, long documents, and many parallel conversations without losing earlier information. GPT-5.1 supports 400,000 tokens.
Category: language
Multilingual
Competence in many natural languages (from a few to over a hundred): understanding, generation, translation, and code-switching within a single conversation. Frontier models support a wide range of languages with comparable quality.
Category: language
Image understanding
Analysing and interpreting the content of images.
Category: vision
OCR
Recognising text within images and documents.
Category: vision
Structured output
Producing data in structured formats such as JSON.
Category: structured_generation
Instruction following
Precisely following instructions contained in the prompt: response format, length, style, constraints (e.g. 'reply in six words'). GPT-5.1 significantly improved this capability compared to GPT-5.
Category: language
Natural conversation
Conducting a conversation with a tone close to human: a warmer voice, empathy in emotional responses, humour, and avoiding the stiff 'AI assistant jargon'. Introduced as a deliberate improvement in GPT-5.1 Instant.
Category: language

Technical architecture

Core Architecture
Training Techniques