
Doubao Seed API 开放了 ByteDance 完整的 Seed 系列,从旗舰级 Seed 2.1 Pro,到专用的 Code Preview,以及高效的 Lite 和 Mini 层级。每个模型都支持深度推理、智能体式规划和仓库级编码,并提供流式传输和共享的 256K 上下文。Atlas Cloud 通过同一个 endpoint 提供全部模型服务,按量付费价格低至每百万 tokens 0.1 美元,并支持 Day-0 access。立即开始构建。
Atlas Cloud 为您提供最新的行业领先创意模型。
Doubao Seed 系列:覆盖前沿推理到高吞吐推理的完整语言模型套件。
| 模型 | 层级 | 最适合 | 上下文 | 最大输出 |
|---|---|---|---|---|
| Doubao Seed 2.0 Pro | 旗舰版 | 复杂推理、智能体、最高难度任务 | 256K | 131K |
| Doubao Seed 2.0 Code Preview | 代码(预览版) | 代码生成、重构、仓库级工作 | 256K | 131K |
| Doubao Seed 2.0 Lite | 均衡型 | 需要以更低成本获得高质量的生产应用 | 256K | 131K |
| Doubao Seed 2.0 Mini | 高效型 | 高并发、延迟敏感、低成本推理 | 256K | 131K |
| Doubao Seed 1.8 | 上一代旗舰版 | 稳定可靠、经过验证的通用推理 | 256K | 131K |
| Doubao Seed 1.6 Flash | 上一代快速版 | 快速响应、实时 UX | 256K | 131K |
| Doubao Seed 1.6 | 上一代基础版 | 高性价比通用任务 | 256K | 65K |
Doubao Seed API 将 ByteDance 完整的 Seed 系列整合到一个兼容 OpenAI 的密钥之下,覆盖旗舰级推理、专用编码模型,以及成本优化的 Lite 和 Mini 档位;它们都共享 256K-token 上下文窗口,并采用按量付费定价,价格低至每百万 tokens $0.1。

旗舰档位 Seed 2.1 Pro 以高端推理、编码和企业级问题解决能力作为整个系列的核心,支持 256K-token 上下文。它能在单次请求中维持长链路、多步骤的思考过程,让复杂的分析和规划任务无需因截断而中断。当最棘手工作负载的正确性比原始成本更重要时,就应选择它,输入价格为每百万 tokens $0.9。

Seed 2.0 Code Preview 专为编码 Agent、仓库理解和日常软件工程进行了调优。同样的 256K 窗口让它能够在编写新函数、重构模块或解释陌生逻辑之前,通读整个代码库。它还可通过 Atlas Cloud Skills 和 MCP Server 接入 AI 编码助手,输入价格为每百万 tokens $0.5。

需要生产级质量,但不想承担旗舰价格?Seed 2.0 Lite 以每百万输入 tokens $0.25 的价格提供接近旗舰的输出,而 Seed 2.0 Mini 则降至仅 $0.1,适合高吞吐量分类、路由和轻量级 Agent 步骤。由于整个系列共享同一接口,你可以将每个请求路由到仍能满足要求的最低成本档位。

该系列中的每个模型都具备 256K-token 上下文窗口,单次调用即可容纳数百页文本。无论输入长文档、多文件仓库,还是冗长的对话历史,它都能保留完整上下文视野。响应会通过标准 server-sent events 逐 token 流式返回,因此即使处理大型任务,你的界面也能保持响应迅速。

定价采用纯按量付费模式,按 token 计费,无需订阅。Prompt 缓存可显著降低重复上下文的成本,其中 Seed 2.0 Mini 的缓存输入价格仅为每百万 tokens $0.02。更长的 prompts 会进入明确公布的更高费率档位,因此当工作负载从原型扩展到生产时,你的支出仍可预测。

应该标准化使用一个档位,还是混合多个档位?一个兼容 OpenAI 的密钥即可访问所有 Doubao Seed 模型,从旗舰 2.1 Pro 到经济型 Mini,再到上一代 1.6 系列。现有 OpenAI SDK 代码只需替换 base-URL 和密钥即可使用,因此你可以在同一个应用中将高难度请求发送给 Pro,将低成本子步骤发送给 Mini。
Feed the identical prompt to three models on the Doubao Seed API and compare how each one turns a single instruction into a working, self-contained HTML page.
Build a complete single-file HTML page — one .html document with all CSS and JavaScript inlined, zero external dependencies, no CDN links, no web fonts, no image files or data URIs. Every visual must be drawn procedurally at runtime with the Canvas 2D API or inline SVG; there must be no bitmaps, photographs, or emoji anywhere. The page is a real-time generative canvas titled something like "Harvest Field — Aerial Geometry Studio," and the entire concept is a god's-eye, top-down view of a cultivated farmland that the user shapes with sliders while a single pixelated combine harvester crawls across it, wiping ordered stripes out of raw crop — the quiet, satisfying moment of order being cut from chaos. Composition and rendering: fill the whole viewport with a regular rectangular grid of rye-like field plots laid out in flat-vector shapes with a subtle isometric / faux-2.5D lift, so each plot reads as a low block, not a flat rectangle. Light the scene as a single directional dusk sun raking in from one corner (warm, low-angle); every plot and the harvester must cast a long soft shadow in the opposite direction to reinforce the geometric, blocky solidity of the field. The palette is a restrained, tasteful earth-tone family: mustard yellow, ochre, olive green, and a single accent of scorched tilled burnt-umber for freshly turned soil — no neon, no pure primaries, keep it high-end and muted. Overall style is flat vector meets isometric micro-relief, precise and architectural. The living behavior: within each plot, render fine drill-seeding rows (thin parallel lines) whose density is adjustable. One pixelated combine harvester — a small blocky sprite drawn from a few rectangles — is the ONLY moving element and the only thing that breaks the grid's symmetry. It advances slowly along either a spiral inward path or a serpentine boustrophedon (回字 / back-and-forth) route across the plots. As it moves it must "erase" the standing crop directly in its swath and leave behind real-time harvest stripes — alternating light and dark cut bands, like a mown lawn or a real combine's header trail — permanently painted onto the field until reset. The harvest is a continuous, self-running animation via requestAnimationFrame; watching it fill the field should feel calm and hypnotic. Interactivity — all controls are sliders (plus a couple of small toggles/buttons) in a clean, minimal side or bottom panel, and every one must respond live without a page reload: a slider for number of field plots (grid resolution, e.g. 3×3 up to 12×12), a slider for drill-seeding row density, a slider for crop hue (shifting the standing crop across the mustard-ochre-olive range), a slider for harvest progress or harvester speed, and a toggle to switch the harvester's path between spiral and boustrophedon. Provide a subtle Reset/Replant button that regenerates the field and restarts the pass. Changing plot count or density should regenerate the layout smoothly; dragging hue should retint in real time. Make it fully responsive so the grid and shadows recompute on window resize, keep the frame rate smooth, and ensure the code is clean and self-contained so it opens correctly by double-clicking the file. This is neither a game nor a dashboard — it is a generative art instrument. Push the procedural geometry, the path animation, and the parameter linkage as far as they go; the quality of the isometric shading, the long dusk shadows, and the crispness of the harvest stripes are what should make this render stand out.
Generated with Doubao Seed 2.1 Pro on Atlas Cloud
Generated with DeepSeek V4 Pro on Atlas Cloud
Generated with Doubao Seed 2.0 Pro on Atlas Cloud
Build a single self-contained HTML file (inline CSS and JavaScript only — absolutely no external libraries, CDNs, frameworks, fonts, or image URLs) that renders an interactive, physically-plausible mechanical pocket-watch movement teardown bench, drawn entirely with SVG and/or Canvas. The page must open and run correctly by double-clicking it in any modern browser, with everything computed and animated in real time. THE SCENE. A top-down view of an exploded brass watch movement laid out on a workshop bench. Render, as living vector art, a mainspring barrel, a stacked gear train (barrel wheel → center wheel → third wheel → fourth wheel → escape wheel), a pallet fork that rocks back and forth, a balance wheel with a spiraling hairspring, and hour/minute/second hands sweeping over a subtle chapter ring. Draw real gear teeth (involute-style flanks, not plain circles), correct meshing at the pitch circles, and visible arbors, jewels, bridges, and screws. The whole assembly is composed concentrically around the movement core; the control panel is tucked into a dark corner. THE MECHANICS (this is the point — make it truly simulate, not fake). Model the gear train with real gear ratios: each wheel's angular velocity equals its neighbor's times the negative ratio of their tooth counts, so meshing wheels visibly counter-rotate at proportional speeds and teeth interlock without overlapping or sliding. The escapement must actually tick: the escape wheel advances one tooth at a time in discrete steps, locked and released by the oscillating pallet fork, which is driven by the balance wheel swinging at a fixed beat. Each release ("tick") produces a crisp step in the escape wheel and an audible tick using the Web Audio API (synthesized, no audio files). The seconds hand advances in beat with the escapement; minute and hour hands derive from it through the gear ratios. Drive the animation with requestAnimationFrame and real elapsed time so timing stays accurate. THE INTERACTIONS (all must respond live). - A draggable winding crown/knob: drag it (mouse or touch) to wind the mainspring — a wind-up gauge fills and the spring coil visibly tightens. Release to let stored energy discharge, powering the train; the movement runs until the spring unwinds, then coasts to a stop. - A "rate" slider that scales the beat/running speed so you can watch it tick slowly or race. - An "explode" slider that continuously separates the stacked components along their axes from fully assembled to a fanned-out exploded diagram, with smooth interpolation, connector guide-lines between separated parts, and everything still rotating correctly at any explosion level. - Hover (or tap) any component to act like a jeweler's loupe: highlight that part, dim the rest, and float a labeled callout with the part's name and a one-line function description. - Buttons: Wind, Release, Reset. THE LOOK. Realistic watchmaking-workshop industrial aesthetic: warm aged brass, distressed steel-grey, deep ink-black background, crisp metallic highlights, finely etched gear teeth and beveled bridges. Simulate a single shaft of window light raking in from the upper-left across the bench so parts cast soft elongated shadows, and let specular highlights travel along the gear faces as they turn. Use gradients, filters, and subtle grain for depth. Clean, legible HUD-style controls and labels in the dark corner. REQUIREMENTS. Everything in one .html file; responsive layout that scales to the viewport; smooth 60fps animation; no placeholder graphics — the movement must be procedurally drawn and genuinely mechanically coupled so that a capable model produces a watch that really winds, meshes, and ticks, while a weak one collapses into static spinning circles. Add brief inline comments naming the gear ratios and the escapement logic.
Generated with Doubao Seed 2.1 Pro on Atlas Cloud
Generated with DeepSeek V4 Pro on Atlas Cloud
Generated with Doubao Seed 2.0 Pro on Atlas Cloud
从编码智能体、仓库级重构,到实时聊天和高吞吐流水线,Doubao Seed API 的各级能力都可通过一个兼容 OpenAI 的 key 接入,嵌入生产技术栈的每一层。
Seed 2.0 Code Preview 针对编码智能体和代码仓库理解进行了优化,可在单次调用中读取 256K-token 上下文。可通过 Atlas Cloud Skills 将其接入 IDE copilot、PR reviewer 和重构 bot。
让 Seed 2.0 Pro 负责规划和工具编排,再将低成本的子步骤委派给 Mini,从而让每次运行的成本可预测。这种拆分适用于研究智能体、工作流自动化,以及需要串联大量调用的流水线。
需要为分类、打标或结构化抽取发起成千上万次调用?Seed 2.0 Mini 面向边缘场景和成本敏感型自动化而构建,让大规模使用时的 token 成本更具优势。
当任务涉及长合同、研究论文或完整知识库时,Seed 2.0 Pro 可在其 256K-token 窗口内进行深度推理。团队可用它对大型文档进行分析、抽取和问答。
一个兼容 OpenAI 的 key 即可访问 Doubao Seed API 的所有层级,从旗舰级 Pro 到高效 Mini。将复杂请求路由到最强模型,将高吞吐任务路由到成本最低的模型,以控制支出。
Seed 2.1 Turbo 和 Seed 1.6 Flash 专为超低延迟、高吞吐和流式响应而设计。即使在高请求量下,它们也能让聊天机器人、实时助手和交互式产品体验保持响应迅速。
了解 Doubao Seed API 与 Atlas Cloud 上领先的推理和编码模型在上下文窗口、最大输出长度以及透明的按 token 计价方面的表现对比。
| 模型 | 提供方 | 上下文窗口 | 最大输出 | 价格(每 1M tokens 输入 / 输出) |
|---|---|---|---|---|
| Doubao Seed 2.1 Pro | ByteDance | 256K | 256K | $0.90 / $4.50 |
| Doubao Seed 2.0 Code Preview | ByteDance | 256K | 128K | $0.50 / $3.00(≤128K prompt),$1.00 / $6.00(>128K) |
| Doubao Seed 2.1 Turbo | ByteDance | 256K | 256K | $0.45 / $2.25 |
| DeepSeek V4 Pro | DeepSeek | 1M | 384K | $1.74 / $3.45 |
| Kimi K2.7 Code | Moonshot AI | 256K | 256K | $0.95 / $4.00 |
| GLM 5.2 | Zhipu AI | 1M | 128K | $1.40 / $4.40 |
| Qwen3.6 Plus | Alibaba Qwen | 1M | 64K | $0.50 / $3.00(≤256K prompt),$2.00 / $6.00(>256K) |
几分钟即可上手 — 按照以下简单步骤,通过 Atlas Cloud 平台集成和部署模型。
在 atlascloud.ai 注册并完成验证。新用户可获得免费额度,用于探索平台和测试模型。
将先进的 Seed 模型与 Atlas Cloud 的 GPU 加速平台相结合,提供无与伦比的性能、可扩展性和开发体验。
低延迟:
GPU 优化推理,实现实时响应。
统一 API:
一次集成,畅用 Seed、GPT、Gemini 和 DeepSeek。
透明定价:
按 Token 计费,支持 Serverless 模式。
开发者体验:
SDK、数据分析、微调工具和模板一应俱全。
可靠性:
99.99% 可用性、RBAC 权限控制、合规日志。
安全与合规:
SOC 2 Type II 认证、HIPAA 合规、美国数据主权。
Doubao Seed API 是 Atlas Cloud 为 ByteDance Seed 系列大语言模型提供的统一端点,面向推理、编码和高并发智能体工作负载而构建。它覆盖旗舰级 Seed 2.1 Pro 和 Seed 2.0 Pro、专用于编码的 Seed 2.0 Code Preview、高性价比的 Lite 和 Mini 层级,以及经过验证的 Seed 1.8 和 1.6 代模型。使用一个兼容 OpenAI 的密钥即可访问所有模型,因此你可以通过单一接口为每项任务匹配合适的层级。
该系列覆盖从前沿推理到高吞吐推理的多种场景。Seed 2.1 Pro 和 Seed 2.0 Pro 适合处理最复杂的推理和智能体任务,Seed 2.1 Turbo 面向低延迟响应,Seed 2.0 Code Preview 专注于软件工程。对成本敏感的流量可使用 Seed 2.0 Lite 和 Mini,而较早的 Seed 1.8 和 1.6 checkpoint 仍可供重视稳定、可预测行为的团队使用。
在 Atlas Cloud 注册,生成一个 API key,然后将现有 OpenAI SDK 指向 Atlas Cloud base URL。计费采用按量付费模式,提供透明的按 token 定价,无需订阅;新的 Seed 版本上线即可 Day-0 访问。现在就开始构建吧。
是的。每个 Doubao 模型都使用标准的 chat-completions 格式,支持 temperature、top_p、frequency 和 presence penalties、最多四个 stop sequences,以及 SSE streaming。现有 OpenAI SDK 代码只需替换 base-URL 和 key 即可使用,无需重写请求逻辑。
该系列所有模型都提供 256K-token(262K)上下文窗口,适合处理大型文档和长对话。最大输出取决于层级:Seed 2.1 模型最高可达 262K tokens,Seed 2.0 层级为 131K,较早的 1.8 和 1.6 checkpoint 为 65K 或 32K。
Atlas Cloud 提供 Python、JavaScript、Swift、Kotlin、Java 和 Dart 的 SDK 访问。你还可以通过 Atlas Cloud Skills 和 MCP Server 将模型连接到 AI 编码助手,这对于基于 Seed 2.0 Code Preview 构建、具备代码仓库感知能力的智能体很实用。
定价采用按 token 付费,无需订阅。标准价格从 Seed 1.6 Flash 每百万 input tokens $0.075、Seed 2.0 Mini $0.1 起,最高到旗舰 Seed 2.1 Pro 的 input $0.9、output $4.5。由于所有层级都在同一个 key 后面,你可以将每个请求路由到满足质量要求的最低成本模型,从而优化开销。
当然可以。单个 Atlas Cloud key 可访问每一代 Seed,因此常见做法是将高难度请求路由到 Seed 2.1 Pro 或 Seed 2.0 Pro,并把高频子步骤交给 Mini 或 Lite。这样既能在关键环节保持质量,又能控制每次运行的成本。
ByteDance 将 Seed 2.1 和 2.0 模型定位为在推理和编码任务上对标美国前沿系统,独立评测也显示其在高难度数学和软件工程任务上具有竞争力。对开发者而言,实际吸引力在于 Doubao Seed API 以低且透明的按 token 费率提供这一等级的能力,价格显著低于大多数西方旗舰模型。
指南、教程与产品动态,助你充分发挥 Atlas Cloud 的价值。