Doubao Seed 2.0 Code

Overview

Doubao Seed 2.0 Code is ByteDance's dedicated coding model. It shipped on February 14, 2026, the same day as the rest of the Seed 2.0 family — ByteDance's launch post describes the release as three general-purpose agent models, Pro, Lite, and Mini, "along with a dedicated Code model."

It has not been superseded. When ByteDance released Seed 2.1 on June 23, 2026, the launch covered two SKUs — Seed 2.1 Pro and Seed 2.1 Turbo — and nothing else. There is no Seed 2.1 Code. Seed 2.0 Code remains the specialist coding model of the Seed line, and the model ByteDance ships inside its own AI IDE.

The design goal is agentic software work rather than autocomplete: the TRAE team frames the model around "more rational decomposition and updating of task plans," which is planning language, not completion language. Unusually for a coding model, it inherits the Seed 2.0 family's multimodal perception — it can read a chart, a screenshot, or a diagram as part of the same request that produces code.

Capabilities

ByteDance's and the TRAE team's stated claims, without embellishment:

• Stronger programming and reasoning abilities than the preceding Seed coding models.
• Agentic task planning: "more rational decomposition and updating of task plans" — the model breaks work down and revises the plan as it executes.
• Multimodal perception: excellent performance in perceiving non-text inputs, carried over from the Seed 2.0 base.
• High-precision text extraction: reading text out of images and documents accurately.
• Chart understanding: interpreting graphs and data visualisations.
• Visual knowledge and reasoning: reasoning over visual content, not merely describing it.

In TRAE, the supported function set is summarised as "reasoning + image understanding" — the model thinks, and it sees.

Technical Specifications

ByteDance publishes little about this model's internals, but Volcano Engine does publish its operational limits. We list only vendor-stated facts.

• Family: Seed 2.0
• Role: dedicated coding model (the only specialist variant in the Seed 2.0 lineup)
• Ark model ID: doubao-seed-2-0-code-preview-260215 — note that it still carries a -preview suffix
• Modalities: text and image input; text output
• Availability: Volcano Engine (Ark) API; TRAE China Edition
• Weights: closed
• Context window: 256K tokens
• Max input: 224K tokens
• Max output: 128K tokens (default 4K)
• Max chain-of-thought: 128K tokens
• Rate limits: 30,000 RPM, 5,000,000 TPM — flagged by Volcano Engine as 非刚性保障 ("not a rigid guarantee")
• Parameter count: Not published by ByteDance.
• Knowledge cutoff: Not published by ByteDance.

Volcano Engine's model list states these as 上下文窗口: 256k / 最大输入: 224k / 最大回答(默认 4k): 128k / 最大思维链: 128k — "context window: 256k / max input: 224k / max output (default 4k): 128k / max chain-of-thought: 128k."

The predecessor endpoint, doubao-seed-code-preview-251028, is marked 即将下线 ("to be taken offline") in the same list, and carries lower limits: a 32K max output and a 32K chain-of-thought, at 5,000 RPM.

Not to be confused with Seed Diffusion

ByteDance Seed separately maintains Seed Diffusion Preview, which its model directory describes as "an experimental diffusion language model specialized for code generation [that] achieves an inference speed of 2,146 token/s." This is a distinct research artefact. It is not a sub-engine, mode, or feature of Doubao Seed 2.0 Code, and the two are routinely conflated in secondary coverage.

Use Cases

• Agentic feature implementation: multi-step work where the model must plan, execute, observe results, and revise the plan.
• Bug diagnosis and repair: tracing a defect across files and proposing a fix.
• Refactoring: restructuring code while preserving behaviour.
• Reading a design and writing the code: because the model accepts images, a mockup, diagram, or screenshot can be part of the prompt.
• Chart and document driven development: extracting figures from a chart or a scanned spec and turning them into logic.
• In-IDE assistance via TRAE: the model's native home, in both SOLO (autonomous) and IDE (interactive) modes.
• UI and screenshot debugging: feeding a rendered screenshot back to the model to reason about what went wrong visually.

Performance / Benchmarks

ByteDance has not published benchmark scores for Doubao Seed 2.0 Code.

That is the honest state of the record. Claims that the model achieves a state-of-the-art result on SWE-bench Verified, or that it ranks first on that leaderboard, appear only in aggregator write-ups and cannot be traced to a ByteDance publication or to a leaderboard entry. Likewise for the frequently quoted "62.7% lower usage cost" figure. We omit all of them rather than launder them into a specification table.

For a sense of where the Seed family sits on public coding leaderboards, the closest primary data point is from the Seed 2.1 launch post, where ByteDance reports that the Seed 2.1 Preview build — published to Arena as Seed-2.1-Pro-Preview — ranks 8th on the third-party Code Arena: Frontend leaderboard with a score of 1539, taking a top-10 position in 5 of 7 frontend subcategories. That is a different model, and we present it only as context.

Limitations

• No published benchmarks. There is no vendor-reported score for this model on any public coding evaluation. You are buying on reputation and your own evals. (The Seed2.0 Model Card benchmarks Seed2.0 Pro, not Code.)
• No published parameter count. The context window, output limit and rate limits are published; the architecture and parameter count are not.
• Still labelled a preview. The current Ark endpoint is doubao-seed-2-0-code-preview-260215, nearly five months after launch.
• Input-length price steps. Cost per token triples between the cheapest and dearest tier, so a long-context agent loop can be far more expensive than a naive per-token estimate suggests.
• No open weights. No self-hosting, no offline use, no local fine-tuning.
• TRAE access is China Edition only. The free in-IDE access ByteDance advertises is for the Chinese edition of TRAE.
• Documentation asymmetry. Chinese-language documentation is substantially more complete than the English material.
• Superseded base model. The model sits on the Seed 2.0 generation while the general-purpose lineup has moved to Seed 2.1. ByteDance has not stated whether a refresh is planned.
• Content policy constraints. As a Chinese-market model, it operates under content rules that differ from those governing US-hosted models.
• Third-party harness compatibility is unconfirmed. Claims of drop-in compatibility with Claude Code, Cursor, Aider, or OpenDevin are not ByteDance statements.

Pricing & Access

ByteDance publishes a rate card, in CNY, on Volcano Engine's 模型价格 (model pricing) page. Pricing for doubao-seed-2.0-code is tiered by input length, where the tier boundaries are given in thousands of tokens.

Under 在线推理（常规）— "online inference (standard)":

Input, cache-hit and output rates are 元/百万token (yuan per million tokens); cache storage is 元/百万token/小时 (yuan per million tokens per hour).

Under 批量推理 ("batch inference"), the published input and output rates are exactly half the standard ones — ¥1.6 / ¥8.0, ¥2.4 / ¥12.0 and ¥4.8 / ¥24.0 across the same three tiers — while cache-hit rates are unchanged and no cache-storage charge applies.

What is confirmed:

• TRAE China Edition — Doubao Seed 2.0 Code is free to use, in both SOLO mode (autonomous agent) and IDE mode (interactive editor).
• Volcano Engine (Ark) — metered API access, at the rates above. This is the only programmatic route.
• Doubao app — ByteDance states that "the Seed 2.0 Pro and Code models have been launched on the Doubao App and TRAE."

Rates change; confirm against the Volcano Engine pricing page before budgeting.

Ecosystem & Tools

• TRAE — ByteDance's AI-native IDE and this model's primary home. Runs Seed 2.0 Code free in the China Edition, in SOLO and IDE modes.
• Volcano Engine (Ark) — the API platform serving the full Seed lineup.
• Seed 2.1 (Doubao) — the current general-purpose flagship, which also handles coding but is not the dedicated code model.
• Seed1.5-VL — ByteDance Seed's standalone vision-language model.
• Seed Diffusion Preview — a separate experimental diffusion language model for code generation, reported at 2,146 token/s.
• Doubao app — carries the Seed 2.0 Pro and Code models for consumers.

Community & Resources

• Seed 2.0 Official Launch — the primary announcement naming the Pro / Lite / Mini / Code family
• TRAE announcement of Doubao-Seed-2.0-Code — the TRAE team's official post, with the capability claims and free-access terms
• ByteDance Seed Models directory — canonical list, including Seed Diffusion Preview
• Volcano Engine model list — the doubao-seed-2-0-code-preview-260215 model ID and its 256K context window
• Volcano Engine model pricing — the tiered CNY rate card
• ByteDance Seed — research team homepage