Volcanic Ark

What is a volcanic ark?

Volcano Ark is a one-stop large-model service and development platform provided by Volcano Engine, offering enterprises and developers capabilities such as model experience, API calls, evaluation, fine-tuning, inference deployment, and application enhancement. Based on the DouBao model, it also integrates mainstream models from the industry to handle tasks in areas such as text generation, deep thinking, visual understanding, speech, images, videos, 3D, and vectorization.

It is not the DouBao app designed for casual conversations among ordinary users, nor is it a single model. Developers can select a model in the Ark console, create an API Key or inference endpoint, and then integrate it into their own customer service, content creation, search, office management, and Agent systems using SDKs, HTTP APIs, CLI tools, or application components.

Core functions

1. Model Plaza and Experience Center

The Model Plaza is used to view the capabilities, context, input types, tool calls, and prices of the DouBao series and models from other suppliers. The Experience Center allows users to test prompts, text, and multimodal inputs directly in the console; after comparing the outputs of different models, they can decide whether to adopt one of them.

Models may have similar names, but their versions, capabilities, and pricing can vary. When selecting a model, it is necessary to check its ID, date of release, context length, as well as whether it supports deep thinking, visual processing, function calls, and structured output; attention should also be paid to any announcements regarding the model’s discontinuation.

2. Chat and Responses API

Ark provides Chat API and Responses API, supporting both streaming and non-streaming output, multiple rounds of messaging, system prompts, tool calls, reasoning parameters, and file input. Its interface style is similar to that of mainstream large-model APIs, making it easy to integrate into web applications, server-side systems, and Agent frameworks.

API requests require that keys be stored on the server, with permissions controlled through IAM, projects, and quotas. Front-end web pages or mobile apps should not contain long-term API keys directly, as this can make it easy for those keys to be extracted, leading to unauthorized usage and associated costs.

3. Deep reasoning, text and code models

The DouBao universal model offers various versions with different levels of performance and cost, such as Pro, Lite, and Mini, and it can handle complex reasoning, long texts, information extraction, writing tasks, as well as Agent-related tasks. The Code model is optimized for code understanding, code generation, and Agentic Coding.

The platform also offers subscription options such as Agent Plan and Coding Plan, which are compatible with CLI and coding tools.

Model rankings and demonstrations cannot replace actual business evaluations. Companies should use their own specialized Chinese terminology, long contexts, structured outputs, tool calls, and secure sample tests, while also keeping track of the accuracy rate, latency, and cost per operation.

4. Image, video, audio, and 3D generation

Ark provides APIs for image generation and editing, video creation, voice-related models, and 3D asset creation. Developers can submit text, images, start and end frames, or other parameters, query the generation tasks asynchronously, and save the results once they are completed.

Multimedia models are typically billed based on images, video length, tokens, characters, or specific specifications; resolution, duration, and whether video input is included all affect the cost. Additionally, the generated content must meet requirements regarding copyright, portrait rights, trademarks, sound licensing, and platform identifiers.

5. Video and image understanding

Visual models can understand the text, objects, scenes, tables, and relationships contained in images, videos, and PDF files; they are used for document question answering, quality inspection, content moderation, and visual positioning. The File API allows large videos, images, or PDF files to be uploaded and preprocessed first, before they can be used in model requests.

Visual models may fail to detect small text, legends, and key frames, and there can also be errors in coordinate determination. Medical images, industrial defects, and safety inspections cannot rely solely on general multi-modal models; specialized data and human review are necessary.

6. Embedding and Vectorization

Text and multimodal vector models can encode text, images, and other content into vectors for use in semantic search, recommendations, clustering, and RAG knowledge bases. The tokenization API allows for an estimation of the number of tokens before a request is made, which helps in managing the context and budget.

The vector spaces generated by different Embedding models are incompatible; therefore, the index must be rebuilt after changing the model. The vector dimension, batch size, normalization method, and distance algorithm also need to match those specified in the vector database configuration.

7. Model evaluation

The platform supports predefined evaluation sets as well as user-defined datasets; evaluations can be initiated from the Model Marketplace, Model Repository, tuning tasks, or machine learning platform services. The inference method can be either online inference or batch inference, and the evaluation results are used to compare different models and their tuned versions.

Online inference is suitable for real-time, small-scale tasks, while batch inference makes use of resources that are not under heavy load to process large datasets where a certain level of delay is acceptable. Evaluations should take into account quality, accuracy, security, latency, stability, and cost, avoiding conclusions based on just a single automated metric.

8. Fine-tuning of SFT and DPO models

Volcano Ark offers supervised fine-tuning for SFT and direct preference optimization for DPO; users upload datasets in the required format, configure the base model and training parameters to generate a fine-tuned model, which can then be deployed as an online inference service. It is suitable for use in scenarios where a fixed output format, industry-specific language, role-based styles, or task-specific enhancements are needed.

Fine-tuning does not equate to injecting real-time knowledge, nor can it replace RAG. Training data needs to be deduplicated, anonymized, authorized, and divided into a validation set;

Errors or biased samples will be learned by the model.

Training, evaluation, and deployment resources may incur costs separately.

9. Optimal Prompt solution

Prompt Optimize provides capabilities for generating, tuning, and managing prompt texts, helping to transform simple requirements into more structured system instructions; it also enables iteration through examples and evaluations. Teams can save different versions of prompts, thereby reducing the need to repeat trial and error for each project.

Automatic optimization may alter constraints or introduce redundancy; before going live, it is necessary to check roles, prohibited items, the output JSON structure, edge cases, and injection protection, as well as to have a version from which recovery is possible.

10. Knowledge base and RAG

The knowledge base plugin allows data to be imported from local sources, object storage, public download links, or Lark documents; it supports formats such as PDF, DOC, DOCX, PPTX, Markdown, TXT, FAQ tables, JSONL, CSV, and XLSX. After the system has parsed, segmented, vectorized, and indexed the data, it becomes possible to search for internal knowledge using model tools.

The Volcano Engine knowledge base also supports mixed semantic and keyword-based searching, rearrangement, scalar filtering, slice modification, traceability of references, as well as the combination of text and images. The search results are used as input for the model, which incurs additional Token usage.

Scans, forms, and images still require sample verification to assess the quality of the parsing.

11. Online search and image processing plugins

Web Search enables the model to access current online information, while image processing tools assist with cropping, recognition, or analysis. The Doubao Assistant tool provides access to common functionalities. Network plugins are charged based on the number of times they are used to retrieve content from various sources, and this usage may also increase the model’s Token count.

The Pro service may also incur minimum consumption fees and expansion costs.

The search results may be outdated, incorrect, or subject to copyright restrictions. The application should indicate the source and date of the information, and it must handle cases where no results are found or when there are conflicts. In medical, financial, and legal contexts, results obtained from online searches cannot be used as final recommendations.

12. Function calls and MCP

Function Calling enables the model to select and enter parameters for business functions based on a predefined schema, which is used to query orders, access databases, or carry out various operations. Ark also supports cloud-based MCP and Remote MCP, allowing external tools to be integrated into the Agent process.

The function parameters generated by the model are not reliable; therefore, the server must verify the type, range, identity, and permissions of these parameters. High-risk operations such as deletion, payment, and sending require additional confirmation. The MCP service should restrict access to networks and keys, as well as prevent prompt injection.

13. Batch inference and context caching

Batch inference is suitable for data annotation, offline summarization, evaluation, and content processing; it improves throughput and reduces costs by utilizing asynchronous tasks. Context caching allows for the reuse of fixed system prompts, long documents, or common prefixes, thereby minimizing repeated computations and associated input costs.

Caching has a validity period and certain rules for hits; dynamic content or user-sensitive data should not be reused across different tenants. Batch tasks require handling of failed attempts, alignment of outputs, task cancellation, as well as considerations regarding data retention periods and cost limits.

14. Stability and cost sharing

The platform offers usage statistics, project-based cost allocation, traffic throttling management, optimization for sudden traffic spikes, and the ability to define custom inference endpoints. The console allows users to view input data, output data, and the total number of tokens for online, experimental, and batch inference tasks, which facilitates cost calculation based on business needs.

Production applications should incorporate features such as timeouts, exponential backoff, concurrent queues, circuit breaking, and alternative models, while also monitoring parameters like TPM, RPM, error rates, and balance. The default rate limiting measures are not sufficient to ensure coverage during peak business loads; it is necessary to conduct load testing in advance and request capacity expansion.

Prices and billing

Volcano Ark charges based on specific models and their capabilities. Text models typically distinguish between input and output tokens, while images, videos, audio, 3D data, Embeddings, knowledge bases, fine-tuning options, batch processing, and plugins all have their own separate rules.

The platform does not have a single fixed monthly fee that covers all capabilities.

As of this verification, the latest reference prices for the DouBao flagship models displayed on the official Volcano Engine website are 6 yuan per million tokens for input and 30 yuan per million tokens for output. These prices apply only to the models shown on the page; they are not valid for Lite, Mini, Code, visual models, older versions, or third-party models.

The New Year 2026 event page also featured offers such as the DouBao 2.0 resource package for 19 yuan, which equates to around 8 million tokens; a package for new customers costing 100 yuan, yielding about 40 million tokens; and Coding Plan Pro at 200 yuan per month. However, some of these products are marked as temporarily unavailable for purchase, and thus cannot be considered as permanent pricing. Companies should refer to the prices listed on the model service page, the activation page in the console, and their order details.

Plugins, knowledge retrieval, and multi-turn tools will result in additional tokens and call counts. It is necessary to estimate the entire business process before use, rather than simply multiplying by the cost per unit of the main model.

Prepaid resource packages, free quotas, bulk discounts, caching discounts, and savings plans also have expiration dates and limitations regarding the models for which they are applicable.

SDK, CLI, and GitHub

Ark provides SDKs for Python, Java, Go, and other languages, as well as an API Explorer and Ark CLI. The Python runtime supports synchronous, asynchronous, and streaming calls; the general Volcano Engine SDK offers AK/SK, STS, role management, timeout settings, retry mechanisms, and proxy configurations.

Volcano Engine has made available on GitHub SDKs for Python, Java, and other languages, as well as agent development kits such as AgentKit and VEADK; some of these are licensed under Apache 2.0. The fact that these SDKs are open source does not mean that the Volcano Ark platform, the weights of the DouBao models, or the hosted inference services are also open source.

The core models and cloud platforms remain proprietary commercial services.

Data security and compliance

Ark provides IAM project permissions, access keys, logging capabilities, quotas, and solutions for ensuring mutual trust. Some custom online services state that request logs are not visible to either the model provider or the platform; however, the way in which data is processed depends on the selected model, the product protocol, and whether data usage has been explicitly authorized.

Platform agreements may include provisions regarding the authorization to use customer data, the regions in which services are provided, and the terms for commercial use. Before uploading training sets, knowledge bases, videos, and facial images, enterprises should verify matters related to personal information, copyright, confidentiality agreements, data export regulations, and industry-specific oversight requirements.

Launching generative AI applications may also require algorithm registration, content labeling, and security assessments.

Volcano Ark Usage Guide

Complete a basic task.

1. Register for Volcano Ark and create an API Key intended solely for the testing environment;
2. Select a model based on input type, context, quality, speed, and price;
3. First, call the Model Plaza and Experience Center to submit a minimal request and check the returned structure;
4. Use the Chat and Responses APIs to test streaming output, parameters, and error responses;
5. Record Tokens, number of calls, latency, error rate, and cost per call;
6. Move the key to the server-side key manager before integrating it into the actual application;

Create reusable professional workflows

1. Different keys and quotas are used for development, testing, and production environments;
2. Representative evaluation sets are established based on the Model Plaza and Experience Center, the Chat and Responses APIs as well as advanced inference capabilities, and text and code models.
3. Set timeout, concurrency, retry, throttling, and budget limits;
4. Perform checks on the output regarding facts, security, format, and sensitive information;
5. Monitor changes in model version, price, latency, and failure rate;
6. Prepare plans for downgrading the model, implementing circuit breaking, and taking manual control;

Which users is it suitable for?

• Developers who need to call the API for DouBao’s text, reasoning, vision, and generation models;
• Companies that wish to carry out selection, evaluation, fine-tuning, and inference on the same platform;
• Form teams to develop intelligent customer service systems, enterprise search tools, as well as systems for content creation and review;
• Applications that require a knowledge base, online searching, functions, and MCP to enhance the agent;
• High-concurrency services that require batch inference, caching, rate limiting, and cost allocation controls;
• Organizations that require domestic cloud services as well as contract and registration-related documentation.

Product advantages

• The DouBao series, along with various other mainstream models, are integrated together;
• Complete set of APIs for text, visuals, images, videos, audio, 3D, and vectors;
• Covers experience, evaluation, SFT/DPO fine-tuning, as well as online and batch inference;
• It provides a knowledge base, networking capabilities, function calls, and an MCP toolchain;
• Context caching, batch tasks, and multi-level models help to optimize costs;
• It offers comprehensive SDKs, CLI tools, IAM functions, monitoring capabilities, and enterprise compliance features.

Restrictions and Precautions

• Volcano Ark offers numerous functions, and its billing components are spread out; the costs related to models, plugins, knowledge bases, and media generation need to be calculated separately.
• The performance shown in the experience center does not reflect the SLAs applicable to production, and the free quotas as well as the promotional prices may expire as well.
• The model may produce factual errors, biases, formatting issues, or incorrect calls to tools;
• Companies must establish mechanisms for evaluation, content security, manual review, permission management, and rollback, as well as pay attention to updates regarding model versions and announcements about their discontinuation.

Frequently Asked Questions

Is Volcano Ark the same as Doubao?

They are different. DouBao is a model as well as a user-facing application brand, while HuoShan FangZhou is a cloud platform that allows enterprises and developers to utilize DouBao and other models for evaluation, fine-tuning, and deployment.

Is Volcano Ark free?

The console may offer a trial experience or a quota for new users, but the official API, advanced customization options, knowledge base, and generation services are charged on a pay-as-you-go basis or through resource packages; the specific details can be found on the model activation page.

How much does the Volcano Ark API cost?

Different models have varying prices. The flagship model shown on the official website currently costs 6 yuan for input and 30 yuan per million tokens for output;

For Lite, Mini, Visual, Video, and plugins, please check their respective prices.

Is it compatible with OpenAI’s interface?

The way in which conversations are initiated while using Ark is similar to that of mainstream Chat interfaces, and it offers Chat, Responses, as well as an official SDK; when performing migration, it is still necessary to verify the Base URL, model ID, parameters, and tool format.

Can the model be fine-tuned?

Yes, it supports SFT and DPO, and allows for the evaluation and deployment of fine-tuned models. Training data, training computations, and online services may incur separate fees.

Is Volcano Ark open source?

The platform and the DouBao model themselves are not open source; the developers have made multi-language SDKs and some Agent development kits available on GitHub, but the license for these SDKs cannot be considered equivalent to the license for the models or cloud services.