ByteDance’s global influence is driven by a sophisticated, layered AI stack. At the base are foundational multimodal models, but its true power comes from industrial-scale systems like Monolith. This recommendation engine, which powers TikTok, is the real-world application of concepts explored in AI product recommendations in e-commerce, turning real-time user behavior into an unparalleled business advantage.
This guide deconstructs the ByteDance AI stack, moving from its core models to the business platforms that productize its power. These tools, like TikTok Symphony, are changing creative workflows by enabling brands to generate content at scale. This new paradigm relies on AI avatar video ads for e-commerce and requires a deep understanding of AI image prompt engineering to achieve high-quality, on-brand results.
Finally, no AI stack of this magnitude can be discussed without its hardware foundation. The massive computational load required for these models is deeply reliant on AI optimizing CUDA code for GPU performance. This hardware layer is also subject to intense geopolitical focus, as seen in the recent China bans on NVIDIA AI GPUs, which directly impacts the future of AI development. This article explains these models, platforms, and their real-world impact for creators, consumers, and companies.
Understanding the ByteDance AI Ecosystem
ByteDance’s AI ecosystem is a layered stack designed to connect foundational research with practical business tools. This structure allows the company to leverage massive-scale data from its consumer apps to train models, which are then productized for creators and enterprises.
Here are the most important components of this stack:
- Doubao: The flagship multimodal model family that powers text, image, and speech applications, processing over 30 trillion tokens daily in China.
- Monolith: The industrial-scale, real-time recommender system that underpins BytePlus Recommend and enables rapid personalization.
- TikTok Symphony: A creative studio that turns text or URLs into ad-ready videos, complete with licensed media and digital avatars.
- Coze: A no-code and low-code platform for building and deploying task-specific AI agents.
Understanding these tiers is key to seeing how ByteDance translates raw AI capability into tangible products for both consumers and businesses. The “Signals to watch” in each section below highlight the key metrics and future developments for each of these platforms.
How to use this guide:
This article is structured in three tiers, moving from core technology to end-user products.
- Tier 1: Foundational Models covers the core AI engines like Doubao and Monolith.
- Tier 2: Business Delivery Platforms explains the enterprise tools, like Symphony and Coze, that are built on top of those models.
- Tier 3: Consumer Creation & Globalization details the apps, like CapCut, that deliver this technology directly to creators and the public.
Start with Tier 1 to understand the core technology, then read Tiers 2 and 3 to see how it is productized.
Tier 1: ByteDance’s Core AI Engines—Inside Doubao LLM and Monolith
Doubao LLM Family (Multimodal Core)
What is Duobao?
Doubao is a family of large models for text, vision, and speech. ByteDance details a visual language model that analyzes charts, code, and scenes. It also features a speech stack that improves text-to-speech realism by capturing accents and emotion. ByteDance details the Doubao visual language model and outlines recent voice and text-to-image upgrades.
Why it Matters
At a national scale, Doubao’s usage signals mainstream adoption, with Volcano Engine reporting over 30 trillion tokens per day by September 2025 (a jump from March) and independent tracking placing the consumer chatbot above 157 million monthly users. Independent reporting places usage at over 30 trillion daily tokens and 157 million monthly users.
Consumer Uses
In daily life, users lean on Doubao for study help, photo understanding, travel planning, and creative tasks. Many of those use cases echo how visual generation and multilingual speech are improving accessibility. For broader context, stronger visual tools lift conversion rates as product imagery improves, a trend that explains how AI avatars shape e-commerce video ads as a downstream effect of better vision models.
Business Uses
Companies embed Doubao via Volcano Engine for chat, search, and analytics. The payoff grows when paired with real-time data and ad tools. Marketing teams align model outputs with channel strategy by pairing generation with audience testing and measurement workflows.
Signals to Watch
- Training efficiency and inference cost curves that determine model reach.
- Upgrades to video understanding, which influence social commerce and safety reviews.
Monolith → BytePlus Recommend (Industrial Recsys)
What is Monolith?
Monolith is ByteDance’s real-time recommendation system. The Monolith architecture combines a collisionless embedding table with expirable embeddings and a serving architecture built for online training. The Monolith paper on arXiv and its real-time learning details describe the architecture and online training approach.
Why it Matters
Unlike most platforms that retrain in batches on static parameters, Monolith is built to learn from live feedback, which is how TikTok-style feeds can respond within minutes to a new trend. BytePlus states that event streams can reflect in recommendations within about 10 seconds of arrival, provided you use the write APIs. Documentation explains the 10-second write-to-serve reflection and the API surface for WriteUserEvents.
Consumer Uses
Viewers see faster cold-start recovery and better niche discovery when the system learns from short sessions. A practical explainer on AI product recommendations in e-commerce shows how similar mechanics improve conversion as inventory and behavior change.
Business Uses
Merchants and publishers can stream catalog updates and user events to improve click-through rate, dwell time, and retention. For adjacent infrastructure decisions, backup strategy and cooling strongly influence the system’s footprint, as scalable AI depends on advanced data center cooling and sustainable server backup practices.
Signals to Watch
- Embedding table memory efficiency and spillover behavior under peak traffic.
- Safety and governance for high-velocity learning, including policy enforcement and audits.
Tier 2: AI for Business—How Symphony, Coze, and Lark Deliver Value
TikTok Symphony Creative Studio (Creative at Scale)
What is Symphony Creative Studio?
Symphony Creative Studio lets teams turn a product description or URL into an ad-ready video, then remix, translate, and add a digital avatar. TikTok’s product page for Symphony Creative Studio outlines the feature set, and Reuters reports the global rollout to advertisers.
Why it Matters
Creative throughput is often the bottleneck in performance marketing. TikTok added access to licensed Getty images and video inside Symphony, which improves legal clarity for brands. Independent coverage documents the Getty integration for Symphony. Practical tactics on scripting and story flow appear in promo video visual storytelling.
Consumer Uses
Viewers get more localized, captioned, and language-matched content because Symphony automates dubbing and translation. That aligns with broader trends we track in voice search and retrieval, where latency and clarity directly shape engagement.
Business Uses
Small teams can ship multi-variant ads for A/B testing without agency timelines. The same pipeline can output organic explainers when teams template scripts, voice, and cuts for rapid iteration.
Signals to Watch
- Disclosure standards for AI-generated creative and avatar usage.
- Asset provenance and rights metadata inside exported files.
Coze (Build Agents Without Heavy Code)
What is Coze?
Coze is ByteDance’s agent platform for building chat and workflow assistants. You can assemble tools, retrieval systems, and prompts without deep coding, or you can extend functionality using the open-source studio. Coze provides a quickstart guide and an open project on GitHub.
Why it Matters
Business tasks are often structured but brittle. Coze turns checklists into agents that call APIs, route messages, and summarize outcomes. For a content team, that might be a brief-builder that reads product pages and drafts scripts. For commerce, it might triage returns and generate resolutions. Our explainer on AI tools for online marketing covers how to align these agents with campaign goals.
Consumer Uses
Individuals can build small automations that summarize class notes, plan trips, or watch price drops. These tasks benefit from the same retrieval and workflow features enterprises use at a larger scale.
Business Uses
Operations teams can prototype agents for support, lead qualification, and internal knowledge search, then promote the best ones to production. For non-video assets, pair agent output with simple image workflows; our overview of AI image enhancement for retail shows how finishing quality affects conversion.
Signals to Watch
- Governance for tool permissions and data scopes as agents call third-party APIs.
- Monitoring and evaluation loops such as success criteria, rubrics, and human overrides.
Lark/Feishu AI Companion (Work OS Layer)
What are Lark and Feishu?
Lark and its China twin, Feishu, embed an organization-wide assistant called AI Companion. This assistant drafts documents, summarizes chats, analyzes PDFs, and helps build simple databases. The feature set and usage patterns are described in Feishu’s AI Companion user manual and its overview of common AI Companion scenarios.
Why they Matter
The most significant productivity win for teams often comes from automating routine work that chews through time: meeting notes, status reports, spreadsheet cleanup, and quick knowledge lookups. Launch coverage of the generative assistant, known as Intelligent Buddy, shows ByteDance positioning a direct productivity layer in Feishu’s work suite, detailed in the Feishu Intelligent Buddy launch report.
Signals to Watch
- Permission scopes for cross-document retrieval and how admins audit agent actions.
- Policy guardrails for summarizing confidential materials in shared spaces.
Volcano Engine / Volcano Ark (Enterprise Delivery & Ecosystem)
What are Volcano Engine and Ark?
Volcano Engine is the enterprise channel for hosting Doubao and related models, with APIs, model management, and compliance features. ByteDance states it is supporting research through a program that offers over 2 trillion tokens for free to universities on the Volcano Ark platform; see the Doubao tokens for academic research announcement.
The company is also expanding compute capacity with large capital expenditure focused on AI infrastructure, according to Reuters reporting on 2025 capex.
Why they Matter
Model access is only useful when delivery and cost controls are stable, which is why enterprises use Volcano Engine to standardize deployment and track usage against budget. For the sustainability side of this expansion, next-generation energy for AI workloads will influence long-term cost and capacity.
Signals to Watch
- Region availability and data residency controls for regulated industries.
- Pricing tiers that differentiate batch inference from interactive sessions.
Tier 3: AI for the Masses—CapCut, SeedEdit, and Global Consumer Apps
CapCut / Jianying (Creator Onramp)
What are CapCut and Jianying?
CapCut exposes ByteDance’s consumer-facing creative stack. The free AI video generator turns a short description into a complete clip with timing, voiceover, and subtitles using the CapCut AI video generator. Jianying is the China counterpart that powers many of the same workflows.
Why they Matter
Lowering the cost of production changes who can speak online. Brands and solo creators use these tools to produce localized variants and to test multiple hooks quickly. For process planning, a focused walkthrough on making an explainer video with AI shows how to script around a single product claim and then scale variants.
Signals to Watch
- Terms of service for generative assets and any watermarking requirements.
- Export quality limits for long-form or high-frame-rate video.
SeedEdit 3.0 & Seedream 4.0 (Pro-Grade Image Models)
What are SeedEdit and Seedream?
SeedEdit 3.0 is a production image-editing model that takes language instructions to retouch, relight, restyle, or add and remove objects while preserving composition. The team describes capabilities and example edits on the SeedEdit product page and in an earlier SeedEdit announcement.
Seedream 4.0 is a next-generation image creation model that unifies text-to-image generation with general-purpose editing in one architecture; the Seedream 4.0 description.
Why they Matter
Consolidating generation and editing removes handoffs between tools and keeps styles consistent across campaigns. ByteDance claims Seedream 4.0 outperforms Google’s Gemini 2.5 Flash Image on an internal benchmark; see outside coverage of the Seedream 4.0 rivalry claim. These results should be treated as ByteDance’s reported figures, not a peer-reviewed comparison.
Business Uses
Retailers and agencies localize catalog imagery, adjust garments, and translate backgrounds for new markets. For downstream performance, evidence that AI background removal drives conversions explains why cleaner product photos increase add-to-cart rates.
Signals to Watch
- Controls for identity likeness, logos, and brand marks in generated images.
- Batch editing speed for high-volume campaigns.
Cici (Global Assistant Outside CN/US)
What is Cici?
Cici is an international assistant that mirrors many Doubao features while operating in markets where Doubao is not distributed. Reporting indicates it blends ByteDance technology with third-party models and is restricted from distribution in China and the United States.
Why it Matters
Cici shows how ByteDance is experimenting with regionalized assistants, which could influence how quickly new features travel from research to everyday use in more countries.
Signals to Watch
- Local disclosure and privacy rules for assistant apps in each market.
- Feature parity with Doubao as multimodal capabilities evolve.
The ByteDance AI Stack: From Models to Business Tools
ByteDance’s AI strategy works because the foundational models and industrial systems are deeply connected. The Doubao LLM provides the core multimodal baseline for understanding, while the Monolith recommendation system turns real-time user behavior into actionable signals. This power is then delivered to businesses and creators through accessible platforms like TikTok Symphony for video, Coze for agent building, and BytePlus for recommendations.
A practical strategy is to match your goal to the ByteDance layer that delivers it. If you need best-in-class personalization for a catalog, study the BytePlus Recommend path. If you need to rapidly scale ad creative, evaluate Symphony and the CapCut onramp. This layered stack, from massive foundational models to no-code business tools, is designed to turn AI capabilities into tangible products.
Key Questions About ByteDance AI Models
Which Models can I Access Today?
Most readers outside China can access the business-facing recommendation stack through BytePlus, and teams can build lightweight agents on Coze without heavy code. An overview of BytePlus Recommend product capabilities describes the discovery experience and common use cases, while Coze provides a documentation hub for assembling agents and workflows. In China, Doubao access routes include the consumer app and enterprise APIs through Volcano Engine.
Is Getty Content in Symphony Licensed?
TikTok states that Symphony includes access to licensed Getty assets for brand use. Getty’s investor newsroom confirms the Getty integration for TikTok advertisers, and TikTok’s product page details Symphony features.
What Does “10-Second Reflection” Mean in Recommendations?
The term refers to low-latency updates between event ingestion and live recommendations, so recent interactions can influence ranking shortly after the signal arrives. BytePlus documents the end-to-end pipeline and APIs that enable this near-real-time behavior in its API integration dashboard.
