Doubao

Doubao (simplified Chinese: 豆包) is an AI assistant and family of large language models developed by ByteDance, the Beijing-based technology company best known internationally as the parent of TikTok and CapCut. Developed by ByteDance's Seed research team, Doubao emerged as China's most widely used AI assistant by monthly active users as of late 2024, reaching 159 million MAU in October 2024 — more than double the user base of its nearest Chinese competitor, Tencent's Yuanbao, at 73 million.

Development and Model Series

ByteDance's AI research arm, operating under the Seed brand, has developed a family of foundation models that underpin Doubao's capabilities. The model lineup includes variants optimised for different deployment contexts.

Doubao Pro series (128K and 256K context variants) are the flagship general-purpose models, designed for complex reasoning, extended document analysis, and multi-turn conversation. The 256K token context window allows processing of approximately 200,000 words in a single session — useful for long document summarisation, code analysis, and research synthesis.

Doubao-Seed-2.0, announced in February 2025, is an advanced model optimised specifically for complex, real-world tasks and "agentic workflows" — tasks that require multi-step planning, tool use, and autonomous task execution rather than single-turn question answering.

In parallel, ByteDance announced UltraMem, a novel memory architecture that reduces AI inference costs by up to 83 percent compared to standard transformer attention mechanisms, reflecting ByteDance's emphasis on cost efficiency for large-scale consumer deployment.

Multimodal Evolution

Doubao's capabilities have expanded substantially beyond text since its initial launch. Version 1.8 of the Doubao Large Model introduced native image understanding. Audio generation was added through a music generation feature (August 2024). Video generation and image understanding followed in late 2024. By 2025, Doubao supported voice interaction, image analysis, video understanding, and code generation within a unified conversational interface.

ByteDance also released Speech-02 through its related MiniMax subsidiary, a text-to-speech model supporting over 30 languages, though Doubao's own voice interaction uses ByteDance's internal speech technology developed for TikTok and Lark (Feishu).

Device Integration

ByteDance has moved to embed Doubao at the operating system level of mobile devices. A prototype developed with ZTE embedded the Doubao LLM into ZTE's Nubia M153 smartphone, running at the OS level to enable the AI to observe the device screen, use installed apps autonomously, pull and organise files, fill forms, and make contextual suggestions. This device-level AI agent approach mirrors strategies pursued by Apple Intelligence and Google Gemini Nano.

Platform and API Access

Doubao is available as a consumer product via the Doubao app on iOS, Android, and web. Developers access the underlying models through ByteDance's Volcano Engine cloud platform (Huoshan Engine API), which provides tiered pricing across model variants. Enterprise clients can access fine-tuned Doubao models for customer service, content generation, and document analysis applications.

See Also

References

References

1. Caixin Global. (2025, February). ByteDance unveils Doubao 2.0 AI model to tackle complex tasks. caixinglobal.com.
2. LLM Reference. (2025). Doubao — ByteDance LLMs. llmreference.com.
3. ByteDance Seed. (2025). Seed team overview. seed.bytedance.com.
4. Winbuzzer. (2025). ByteDance and ZTE unveil agentic AI smartphone prototype. winbuzzer.com.
5. Scientific American. (2025). ByteDance launches Doubao real-time AI voice assistant for phones. scientificamerican.com.