AI Policy Wiki
Dashboard

Doubao

medium confidence · updated 2026-06-06

ByteDance's Doubao (豆包) model family — consumer chatbot backbone with Pro, Lite, Pro-1.5, Vision, and reasoning variants. Largely closed; free to consumers; distributed via Douyin / TikTok scale.

Doubao (豆包) is ByteDance's consumer-facing large language model family, distributed at scale through the Doubao app in China, the Cici app overseas, and AI features inside Douyin and TikTok. The family is closed-weight and API-only, free to consumers since launch, and domestic reporting frequently describes it as the largest-MAU Chinese AI chatbot (a claim of medium confidence, as user-count figures vary by source). Public technical disclosure is thinner than for peer Chinese developers such as DeepSeek, Alibaba, or Moonshot.

FieldValue
Developer[[bytedance-doubaoByteDance]]
Model family nameDoubao (豆包) / Seed (internal research series)
Launched2023 (chatbot); Pro-tier named releases 2024–2025
Primary deploymentDoubao app (China); Cici (overseas); Douyin / TikTok AI features; Lark
LicenseClosed-weight; API-only
Consumer pricingFree to consumers since launch (Source: China and the US Are Running Different AI Races)

Model variants

The model line has progressed through several named variants, released on an iteration cadence that follows ByteDance's consumer-product culture rather than the paper-and-benchmark release pattern of research labs. Naming is fluid, and exact naming and dates carry medium confidence. The variants documented through 2025 are:

ModelNotes
Doubao ProFlagship general-purpose
Doubao LiteEfficiency tier
Doubao Pro-1.52024 refresh
Doubao-VisionMultimodal variant
Doubao-Music / Doubao-VideoSpecialist generation models
Seed-1.5 / Seed-ThinkingInternal research series; arXiv papers

ByteDance's internal research series, Seed, publishes occasional arXiv papers (Seed-1.5, Seed-Thinking) that document underlying techniques, sometimes describing methods without releasing weights.

Capabilities

Doubao supports vision, audio (including music generation), and video understanding. Its distribution is integrated into Douyin (reported at more than 700M DAU in China) and TikTok, giving it more user-facing exposure than any other Chinese model. The lineup has reportedly added reasoning-mode variants following the DeepSeek-R1 and o1 pattern; this is reported at medium confidence.

Open-weight status

Doubao is closed and API-only. ByteDance has not released open weights for the flagship Doubao family, in contrast to Qwen3 (Apache 2.0), DeepSeek-V3 (fully open), and GLM-4-9B (partially open). The Seed research arXiv papers sometimes describe techniques without releasing the corresponding weights.

Deployment and product split

Doubao runs as a split domestic and overseas product. In China it is offered through the Doubao app, Douyin integration, and Lark (enterprise), hosted on PRC infrastructure, registered under China — Interim Measures for the Management of Generative AI Services, and content-aligned per CAC rules. Overseas it is offered through Cici (a standalone app), Coze (an agent-builder platform), and AI features inside TikTok, hosted globally and subject to US, EU, and UK regulation.

This dual-stack commercial architecture is not present in other Chinese frontier models. It mirrors ByteDance's broader TikTok and Douyin legal-separation pattern and bears on AI sovereignty debates as a commercial instance of a domestic/overseas dual stack.

Pricing and distribution

Doubao has been free to consumers since launch, and ByteDance reportedly cut API prices in 2024 during a round of Chinese-market price wars. This follows the consumer-AI pricing pattern described in China and the US Are Running Different AI Races, in which consumer AI is free (Baidu, ByteDance) and monetization comes through API, cloud, and advertising rather than ChatGPT-Plus-style subscriptions.

Reported ByteDance AI capital expenditure of roughly $21B in 2024–2025 (Source: China and the US Are Running Different AI Races) places single-firm Doubao spend in the range of US hyperscalers, though still below frontier-lab training commitments per China and the US Are Running Different AI Races.

Regulatory environment

In China, Doubao is registered as a public generative-AI service under China — Interim Measures for the Management of Generative AI Services. The China — Provisions on the Administration of Deep Synthesis Internet Information Services apply to Doubao's music, image, and video surfaces and to Douyin's synthetic-media features, and the China — Internet Information Service Algorithmic Recommendation Management Provisions treat Douyin as a principal implicit target; Douyin's use of Doubao-class models for recommendation and synthesis is part of why. Overseas, Cici and TikTok AI features are subject to the EU AI Act, US state law, and UK regulation, making Doubao the most Western-regulator-exposed Chinese model.

Doubao's consumer scale contrasts with thin public safety-case disclosure, an instance of the deployment-versus-frontier-safety asymmetry described in China and the US Are Running Different AI Races.

Comparison to peer models

ModelDeveloperOpen-weightDistribution strength
DoubaoByteDanceClosedTikTok/Douyin captive scale; bifurcated global stack
[[ernie-4ERNIE 4]]BaiduClosedBaidu search; robotaxi; longest tenure
[[glm-4GLM-4]]ZhipuPartial (9B open)Enterprise API; Tsinghua pedigree
[[qwen3Qwen3]]AlibabaApache 2.0Alibaba Cloud + permissive license
[[deepseek-v3DeepSeek-V3]]DeepSeekFully openResearch reputation + efficiency
[[kimi-k2Kimi K2]]MoonshotResearch-only licenseLong-context agentic focus

Doubao and ERNIE cluster as closed, consumer-distributed, free-to-user Chinese models. Qwen and DeepSeek cluster as open-weight frontier research, with GLM in between. In the framing of US-China AI competition, Doubao illustrates the Chinese "consumer free, monetize elsewhere" pricing pattern and the single-firm-capex scale of ByteDance.

Relationships

Sources

Referenced via ByteDance / Doubao and China and the US Are Running Different AI Races. Confidence is medium: ByteDance publishes fewer frontier technical disclosures than peer Chinese labs, and most claims rely on secondary reporting and Seed arXiv papers.