Doubao

Doubao (豆包) is a family of large language models developed by ByteDance Seed, with the consumer Doubao app launched in August 2023 and the Doubao 1.5 Pro frontier model released in January 2025. The closed-weights production line powers the Doubao chatbot app, integrations into Douyin and Toutiao, and enterprise API access through Volcano Engine; the multimodal Doubao 1.5 Vision Pro and the Seed1.5-VL technical-report variant extend the line into vision tasks. As of April 2026, Doubao is the most-used native Chinese consumer AI assistant by monthly active users, surpassing DeepSeek through 2025 and ranked among the top consumer AI applications globally, with the Doubao Seed 2.0 release in February 2026 establishing the line at the frontier capability tier.

At a glance

• Lab: ByteDance Seed
• Released: Doubao consumer app (August 2023); Doubao Pro (May 2024) and consumer-LLM price-war pricing; Doubao 1.5 Pro and 1.5-Vision-Pro (January 22, 2025); Seed1.5-VL technical report (May 13, 2025); Doubao Seed 2.0 family (February 14, 2026)
• Modality: Text and multimodal (vision, audio in dedicated variants); video and image generation through sibling Seedance and Seedream lines
• Open weights: No for the production Doubao variants. Selected ByteDance Seed open-weights releases include Seed1.5-VL, Seed-OSS, and UI-TARS for agentic-interface use cases.
• Context window: Up to 256,000 tokens for Doubao 1.5 Pro variants; varies by sub-model
• Pricing: Doubao 1.5 Pro at approximately $0.11 per million input tokens and $0.275 per million output tokens; cached input at $0.022 per million tokens. Doubao Seed 2.0 Pro at approximately $0.47 input and $2.37 output per million tokens. Pricing is among the lowest in the Chinese frontier-model tier and was a primary driver of the Chinese LLM price war from May 2024.
• Distribution channels: Doubao consumer app (web, iOS, Android), embedded Doubao features in Douyin and Toutiao, Volcano Engine for enterprise API access, sibling international Cici and Dola consumer apps

Origins

The Doubao consumer chatbot launched in August 2023 as ByteDance's entry into the Chinese consumer LLM market. The first underlying models were ByteDance internal language models with capability comparable to mid-tier Chinese-market peers; the consumer brand and rapid distribution scaling through Douyin and Toutiao were the line's distinguishing positions in the first eighteen months.

The May 2024 release of the Doubao Pro enterprise-API tier through Volcano Engine reset Chinese LLM pricing. Doubao Pro was offered at approximately 0.0008 yuan per 1,000-token prompt, roughly 99.8% below OpenAI's contemporaneous GPT-4 rates and 99% below prevailing Chinese enterprise-LLM rates from Baidu, Alibaba, and Tencent. The pricing was the explicit trigger for the Chinese LLM price war, with Baidu, Alibaba, Tencent, iFlytek, and others responding within days. The price war reshaped Chinese enterprise-LLM economics through 2024 and into 2025.

The January 22, 2025 release of Doubao 1.5 Pro was ByteDance's first competitive frontier-tier release, using a sparse mixture-of-experts architecture with 7 active experts and 20 billion activated parameters per token. ByteDance reported that Doubao 1.5 Pro outperformed GPT-4o, Claude 3.5 Sonnet, DeepSeek-V3, and Llama 3.1 405B on multiple benchmarks including AIME and the company's SuperGPQA evaluation.

Doubao 1.5 Vision Pro and Doubao 1.5 Real-time Voice Pro accompanied the release, extending the family into multimodal vision and voice. The vision variant was further detailed in the May 13, 2025 Seed1.5-VL technical report, which documented state-of-the-art results on 38 of 60 public vision-language benchmarks and 14 of 19 video-understanding benchmarks. Seed1.5-VL paired a 532-million-parameter vision encoder with a 20-billion-active-parameter MoE language model.

The most consequential release came on February 14, 2026, with the Doubao Seed 2.0 family, ByteDance's first release after Wu Yonghui (former Google Fellow) joined Seed as research head in February 2025. Doubao Seed 2.0 shipped in four variants (Pro, Lite, Mini, and Code), with the Pro flagship reported as a 1-trillion-parameter Gemini-style multimodal model. Reported benchmark performance for Doubao Seed 2.0 Pro included 98.3 on AIME 2025, 88.9 on GPQA Diamond, 76.5 on SWE-Bench Verified, and a 3020 Codeforces rating, placing the model in the leading group of Chinese frontier releases.

The Doubao consumer app accumulated user counts faster than peer Chinese LLM consumer products through 2025. By August 2025, Doubao reached 157 million monthly active users in China, surpassing DeepSeek's chat product to claim the top position among native Chinese AI applications. By the end of 2025, Doubao had 226 million monthly active users, and a16z ranked Doubao as the fourth-most-popular generative AI app worldwide alongside ChatGPT and Gemini.

Capabilities

Doubao 1.5 Pro and Doubao Seed 2.0 cover general-purpose Chinese and English text generation, multi-turn dialogue, code generation, mathematical reasoning, and document analysis, with multimodal variants extending into vision and voice.

The Doubao 1.5 Pro mixture-of-experts configuration uses a sparse activation pattern of 7 experts and 20 billion activated parameters per token. The architecture was designed for inference-cost minimization at production scale: ByteDance reported 50% gross margins on Volcano Engine API revenue at launch. The deliberately low active-parameter count relative to peer Chinese MoE models (DeepSeek V3's 37 billion, Kimi K2's 32 billion) reduced per-token compute and made the aggressive price positioning sustainable.

Chinese-language capability is the line's strongest axis. Doubao 1.5 Pro reportedly leads peers on Chinese-domain benchmarks including SuperGPQA and on Chinese-language portions of standard multilingual benchmarks. English-language performance is competitive but does not consistently lead.

Doubao 1.5 Vision Pro (Seed1.5-VL) handles image understanding, document OCR, chart analysis, GUI control, and video understanding. Seed1.5-VL documented state-of-the-art performance on 38 of 60 public vision-language benchmarks, leading on 14 of 19 video benchmarks. In agent-centric tasks including GUI control and gameplay, the model outperformed OpenAI's CUA and Anthropic's Claude 3.7. Doubao 1.5 Real-time Voice Pro provides voice-based interaction with low latency for the consumer app and enterprise deployments.

The Doubao Seed 2.0 release in February 2026 substantially raised the line's ceiling on reasoning and coding benchmarks, placing it in the leading group of Chinese frontier releases alongside DeepSeek V4 and the latest Qwen and Hunyuan generations. Doubao Seed 2.0 Code is the dedicated coding variant; Lite and Mini are smaller-context, lower-cost siblings.

The consumer app integrates the underlying models with voice input, image input, document upload, real-time search, and a tooling layer. Daily token consumption across the product surface reached 120 trillion tokens by late 2025, a thousand-fold increase over two years and one of the largest reported consumer-LLM throughput figures globally.

Benchmarks and standing

Doubao's benchmark profile has shifted substantially across the 1.5 to 2.0 generations, reflecting ByteDance's investment in frontier-tier capability through 2025 and 2026.

At the January 2025 release, Doubao 1.5 Pro reported leading scores on multiple Chinese-language and bilingual benchmarks. On SuperGPQA (ByteDance's Chinese-language graduate-knowledge evaluation), Doubao 1.5 Pro ranked first among chat models with 55.09% accuracy. On AIME 2025, the model reportedly outperformed DeepSeek-V3 and Llama 3.1 405B. ByteDance's claim of matching or exceeding GPT-4o and Claude 3.5 Sonnet was supported by selected internal-benchmark results but not universally validated by independent evaluations.

On vision-language benchmarks, the Seed1.5-VL technical report documented state-of-the-art results on 38 of 60 public benchmarks, including leadership on 14 of 19 video-understanding benchmarks and on agent-centric GUI control and gameplay tasks.

The Doubao Seed 2.0 Pro release in February 2026 raised the line's frontier-tier positioning. Reported scores of 98.3 on AIME 2025 and 88.9 on GPQA Diamond placed the model in the leading group at the release date. SWE-Bench Verified at 76.5 was competitive with the leading frontier-tier coding models. The 3020 Codeforces rating was the highest among Chinese-developed models at the release.

The Artificial Analysis Intelligence Index has not consistently published Doubao scores at the cadence of US frontier-lab releases, reflecting the closed-weights and Chinese-market focus. On LMArena, the line has appeared at competitive positions among Chinese closed-weights frontier models without consistently leading.

The line's distinguishing benchmark axis is Chinese-language performance combined with cost-per-token: Doubao 1.5 Pro's price-performance ratio at the January 2025 release was approximately 5 times below DeepSeek and 200 times below OpenAI's o1 on per-token rates, with comparable benchmark coverage. Benchmark leadership rotates rapidly; Doubao's positions are representative of the January 2025 through February 2026 release period.

Access and pricing

Doubao distribution combines mass-market consumer-app access with enterprise API tiers and closed-weights gating across the production line.

The consumer Doubao app at doubao.com provides free public access to Doubao 1.5 Pro and successor models on web, iOS, and Android. Voice and image inputs are supported. The international consumer apps Cici and Dola, released initially in August 2023, provide localized international access.

Enterprise API access runs through Volcano Engine at volcengine.com/product/doubao. Per-token pricing for Doubao 1.5 Pro is approximately $0.11 per million input tokens, $0.275 per million output tokens, and $0.022 per million cached input tokens. Doubao Seed 2.0 Pro pricing is approximately $0.47 input and $2.37 output per million tokens, reflecting the higher capability tier. The Lite, Mini, and Code variants are offered at substantially lower per-token rates.

The pricing aggression that triggered the Chinese LLM price war in May 2024 has continued to characterize the line. ByteDance has consistently positioned Doubao API pricing below peer Chinese frontier-model rates and substantially below US frontier-lab equivalents.

There is no first-party open-weights distribution of the production variants. Open-weights releases from ByteDance Seed (Seed1.5-VL technical-report variants, Seed-OSS, UI-TARS, selected Seedance variants) are distributed separately through ByteDance-Seed on GitHub under research-and-commercial licenses appropriate to each release.

Comparison

Direct competitors to Doubao across the 2024 to 2026 active period:

• DeepSeek V3 and DeepSeek V4 (DeepSeek). The principal Chinese open-weights peer at the frontier tier. DeepSeek's open-weights distribution and consumer chat product compete directly with Doubao's consumer and enterprise offerings. DeepSeek surpassed Doubao briefly in early 2025 following R1; Doubao reclaimed the consumer-app top position by August 2025. The two are the most-used Chinese consumer AI products.
• Ernie (Baidu). Baidu's closed-weights and partially open-weights line. ERNIE 4.5 and ERNIE X1 cover the same general capability profile as Doubao 1.5 Pro, with ERNIE leading on selected Chinese-language benchmarks. Doubao's distribution scale through Douyin and Toutiao has produced substantially larger consumer user counts than Ernie Bot.
• Hunyuan (Tencent). Tencent's parallel Chinese frontier line with WeChat distribution. Doubao's advantage is the consumer-app distribution scale; Hunyuan's advantage is WeChat integration and Tencent enterprise relationships.
• Qwen 3 (Alibaba Qwen). The major Chinese open-weights frontier line. Qwen 3's variant breadth and multimodal coverage exceed Doubao's open-weights footprint; Doubao's closed-weights production tier and consumer-app distribution exceed Qwen 3's consumer-product penetration.
• GPT-4o, GPT-5 (OpenAI), Claude 3.5 Sonnet (Anthropic). US closed-weights frontier comparators that Doubao 1.5 Pro and Doubao Seed 2.0 Pro target on benchmarks. The two are not comparable on consumer-market deployment given Chinese and US regulatory separation. Doubao's price-per-token positioning is approximately 50 to 200 times below US frontier-lab rates at comparable releases.

Doubao's distinctive position: the largest Chinese consumer-app distribution by monthly active users, the lowest production-API pricing in the frontier tier (driver of the Chinese LLM price war from May 2024), the multimodal Seed1.5-VL contribution to vision-language and agent-centric benchmarks, and the Doubao Seed 2.0 February 2026 release that established the line at frontier-tier capability.

Outlook

Open questions for Doubao over the next 6 to 18 months:

• Doubao Seed 2.0 enterprise adoption. The February 2026 release positioned Doubao at frontier-tier benchmarks for the first time. Whether the Pro flagship achieves enterprise-API adoption beyond Chinese-market deployments through Volcano Engine's international presence is an open commercial question.
• Pricing economics post-price-war. Whether ByteDance maintains the pricing posture as Doubao Seed 2.0 raises model-line compute cost, or shifts toward higher rates that match capability tiers, will affect both consumer-app economics and competitive dynamics with DeepSeek, Ernie, and Hunyuan.
• Open-weights direction. ByteDance Seed has released selected open-weights research models but has not open-sourced the production Doubao variants. Whether Doubao Seed 2.0 or successors follow the open-weights direction taken by DeepSeek, Qwen, and Ernie 4.5 will affect ByteDance's positioning in the open-weights ecosystem.
• International consumer expansion. The Cici and Dola international apps have not achieved comparable user counts to Doubao. Whether the line achieves international consumer scale, particularly in Southeast Asia and emerging markets, is a strategic question.
• Multimodal and agentic depth. The Seed1.5-VL agent-centric capability and Doubao Seed 2.0 Pro coding strength position the line for agentic deployments. Whether ByteDance pursues frontier agentic-AI deployment at the level of Anthropic Claude Code or OpenAI Codex will shape the line's long-term capability emphasis.

Sources

• Seed Models on the ByteDance Seed Team site. Official documentation of the Doubao Seed 2.0 family and Seed1.5-VL technical context.
• GitHub: ByteDance-Seed organization. Open-weights research-model repositories for Seed-OSS, Seed1.5-VL, UI-TARS, and selected Seedance variants.
• Volcano Engine: Doubao product page. Enterprise-API documentation, pricing, and product variants.
• Wikipedia: Doubao. Consumer-app launch date, product history, and market position.
• Seed1.5-VL Technical Report (arXiv). Technical report for the multimodal vision-language variant, May 2025.
• SCMP: ByteDance launches low-cost Doubao AI models for enterprises. Coverage of the May 2024 enterprise-API launch and the Chinese LLM price war trigger.
• MarkTechPost: ByteDance AI Introduces Doubao 1.5 Pro. January 2025 Doubao 1.5 Pro release coverage with architectural and benchmark detail.
• QuestMobile via aibase: Doubao surpasses DeepSeek in monthly active users. August 2025 user-count data positioning Doubao at the top of Chinese consumer AI applications.
• a16z: Top 50 Generative AI Apps. Industry-coverage source positioning Doubao among the top global generative AI apps by reach.