Doubao Seed Code

doubaodoubao/doubao-seed-code

Doubao Seed Code — code generation and code understanding model.

36Kr — Doubao Seed Code launch coverage (Nov 2025) ↗Volcano Engine — Doubao models (API access) ↗Volcano Engine docs — Doubao-Seed-Code model page ↗

Doubao Seed Code is ByteDance's first dedicated programming model, released on November 11, 2025 by Volcengine (the cloud and AI service arm of ByteDance). It was purpose-built for Agentic Coding tasks — code generation, debugging, refactoring, code review, and unit test creation across 200+ programming languages. At launch it was the first domestic (Chinese) coding model to support visual understanding, enabling UI-to-code workflows from screenshots or design mockups.

At launch it achieved best-in-class scores among Chinese models on SWE-Bench-Verified (powering the Trae IDE combo at SOTA), HumanEval, and Aider benchmarks, outperforming DeepSeek-V3.1, Kimi-K2, and GLM-4.6 — second only to Claude Sonnet 4.5 among all AI programming models. For TheRouter operators, this model fills the 'premium Chinese coding' slot alongside the later Seed 2.0 Code; its pricing via TheRouter ($0.1667/$1.1111 per million tokens) makes it the most affordable Anthropic-API-compatible Chinese coding endpoint. The model supports a native 256K context on Volcengine Ark (exposed as 131K via TheRouter) and is fully compatible with Claude Code, Trae, Cursor, and veCLI tools.

Best for

• Chinese-first agentic coding workflows (Claude Code, Cursor, Aider, Continue.dev) needing a budget alternative to Claude Sonnet 4.5 with strong SWE-Bench performance and Anthropic API compatibility.
• UI-to-code and front-end prototyping from screenshots or design mockups — the model's visual understanding capability is unique among Chinese coding models at its price tier.
• Mixed Chinese/English repository contexts up to 131K tokens, where the model's training on Chinese-first codebases (Trae, internal ByteDance tools) gives an edge over English-only coding models.

Reach for something else if

• Pure English agentic or non-coding reasoning tasks — Seed 2.0 Code, Claude Sonnet 4.6, or DeepSeek-V3.2 deliver stronger measured throughput and broader capability at similar or lower cost.
• Workloads requiring open-source weights or self-hosting — Seed Code is API-only via Volcengine; use Qwen3-Coder or Seed-Coder (open-source ByteDance weights) for local deployment.

Context Length

131K

Max Output

33K

Input Priceper 1M tokens

$0.1944/ 1M tokens

Output Priceper 1M tokens

$1.30/ 1M tokens

Modalities

text→text

Pricing Breakdown

Type	Rate
Input	$0.1944 / 1M tokens
Output	$1.30 / 1M tokens

Supported Parameters

temperaturemax_tokenstop_ptoolstool_choiceresponse_formatstop

Specifications

Release date	2025-11-11eu.36kr.com ↗	verified
Upstream model ID	doubao-seed-code-preview-251028 (Volcengine Ark)blog.laozhang.ai ↗	verified
Native context window	256,000 tokens (Volcengine Ark); 131,072 tokens via TheRouterblog.laozhang.ai ↗	verified
Architecture	Doubao base model family, fine-tuned for programming tasksdev.to ↗	to verify
Visual understanding	Supported — first Chinese coding model with image-to-code capabilityeu.36kr.com ↗	verified
API compatibility	Anthropic API-native (Claude Code drop-in); also OpenAI-compatibleeu.36kr.com ↗	verified
Training cutoff	Not publicly disclosed	unknown
License	Proprietary (API-only — closed weights)www.volcengine.com ↗	verified
Supported tools	Claude Code, Trae (Chinese IDE), veCLI, Cursor, Volcano Engine AI Experience Centereu.36kr.com ↗	verified

Benchmarks

Benchmark	Score	Source
SWE-Bench-Verified Topped the SWE-Bench-Verified leaderboard at launch when paired with Trae IDE (ByteDance's AI-native IDE).	SOTA (Trae + Seed Code combo)	eu.36kr.com ↗
HumanEval ByteDance reported performance above DeepSeek-V3.1, Kimi-K2, and GLM-4.6 across major coding benchmarks at launch; exact numeric scores not published separately.	Not publicly disclosed (pre-Seed 2.0)	eu.36kr.com ↗
Aider Trae + Seed Code scored among the top on Aider-polyglot coding benchmarks as reported at launch. See Seed 2.0 Code page for the successor model's numeric results.	Not publicly disclosed (pre-Seed 2.0)	eu.36kr.com ↗
Comparative ranking (launch) At launch, Doubao Seed Code outperformed all other Chinese coding models on major benchmarks and ranked second globally only to Claude Sonnet 4.5 among AI programming models.	> DeepSeek-V3.1, Kimi-K2, GLM-4.6; < Claude Sonnet 4.5	eu.36kr.com ↗

API Usage Examples

Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.

cURL

curl https://api.therouter.ai/v1/chat/completions   -H "Content-Type: application/json"   -H "Authorization: Bearer $THE_ROUTER_API_KEY"   -d '{
    "model": "doubao/doubao-seed-code",
    "messages": [
      {"role": "user", "content": "Summarize the key points from this input."}
    ]
  }'

API guide

Full API reference →

Chat completion

cURL

curl https://api.therouter.ai/v1/chat/completions \
  -H "Authorization: Bearer $THEROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"doubao/doubao-seed-code","messages":[{"role":"user","content":"用 Python 写一个带类型注解的二分查找函数。"}]}'

More from doubao

doubao/doubao-seed-2-0-codeSuperseding model — Seed 2.0 architecture with diffusion-based code generation, higher benchmarks (87.8 LiveCodeBench, 76.5 SWE-Bench), and 256K native context at slightly higher cost.doubao/doubao-seed-2-0-proFlagship generalist with identical coding capabilities plus stronger multimodal and long-horizon reasoning — useful when the task mix includes both code and non-code work.doubao/doubao-seed-1-8Predecessor general-purpose model (December 2025) — the base that Seed Code's architecture was derived from.

Similar models

Cross-provider sibling models

zhipu/glm-4.7

GLM 4.7 Coding — Zhipu's answer to Doubao Seed Code with similar pricing and Chinese-first coding focus. Comparable benchmark range but different tool ecosystem (CodeGeeX).

moonshot/kimi-k2

Competing Chinese coding model from Moonshot AI with strong multilingual code understanding. Slightly lower SWE-Bench than Seed Code but competitive pricing.

deepseek/deepseek-v3.1

DeepSeek's strongest non-reasoning model (pre-V3.2). Weaker than Seed Code on SWE-Bench/HumanEval at launch but more expensive on output tokens. Note: V3.2 supersedes this model.

News & changes

2025-11-11

ByteDance releases Doubao Seed Code — first coding model, SOTA on SWE-Bench

Volcengine launched Doubao Seed Code, ByteDance's first dedicated programming model. Outperformed all Chinese competitors on SWE-Bench-Verified, HumanEval, and Aider benchmarks. Only Claude Sonnet 4.5 ranked higher globally. The model introduced visual code generation from screenshots and native Claude Code compatibility — the first for any Chinese model.

re-authored by TheRouter36Kr ↗

2025-11-11

Doubao Seed Code integrated into Trae IDE

ByteDance's AI-native Chinese IDE Trae shipped with Seed Code pre-integrated, achieving the top SWE-Bench-Verified score as a combined model + tool stack. The integration lets developers use Seed Code directly within the Chinese IDE for agentic code generation.

re-authored by TheRouter36Kr ↗

Frequently asked

What is the difference between Doubao Seed Code and Doubao Seed 2.0 Code?

Doubao Seed Code (November 2025) is ByteDance's first-generation programming model, built on the original Doubao architecture. Doubao Seed 2.0 Code (February 2026) is the successor built on the newer Seed 2.0 architecture with MoE, diffusion-based ultra-fast generation, higher benchmarks (87.8 LiveCodeBench v6 vs unquantified first-gen), and native 256K context. For new projects, Seed 2.0 Code is the recommended choice. The original Seed Code remains available via TheRouter for backward compatibility with workflows that were tuned on the first-generation model.

re-authored by TheRoutereu.36kr.com ↗

Can I use Doubao Seed Code with Claude Code or Cursor?

Yes. Doubao Seed Code is natively compatible with the Anthropic Messages API, making it a direct drop-in replacement for Claude Code. It also works with Cursor, Aider, Continue.dev, and any OpenAI-compatible tool through api.therouter.ai/v1. For visual code generation (screenshots to code), the model supports image input through the standard chat completions interface.

re-authored by TheRouter

What programming languages does Doubao Seed Code support?

Doubao Seed Code supports 200+ programming languages including Python, JavaScript/TypeScript, Java, Go, C/C++, Rust, PHP, Ruby, Swift, Kotlin, and many others. It excels particularly at Chinese-dominant codebases and mixed Chinese-English projects due to the native Chinese training corpus. For front-end development, it has specialized strengths in Vue, React, and CSS with screenshot-to-code capability.

re-authored by TheRouter

Fact ledger — every claim on this page traces here

source	URL	retrieved
Release date	eu.36kr.com ↗	2026-05-26	verified
Upstream model ID	blog.laozhang.ai ↗	2026-05-26	verified
Native context window	blog.laozhang.ai ↗	2026-05-26	verified
Architecture	dev.to ↗	2026-05-26	to verify
Visual understanding	eu.36kr.com ↗	2026-05-26	verified
API compatibility	eu.36kr.com ↗	2026-05-26	verified
Training cutoff	—	—	unknown
License	www.volcengine.com ↗	2026-05-26	verified
Supported tools	eu.36kr.com ↗	2026-05-26	verified
SWE-Bench-Verified	eu.36kr.com ↗	2026-05-26	verified
HumanEval	eu.36kr.com ↗	2026-05-26	to verify
Aider	eu.36kr.com ↗	2026-05-26	to verify
Comparative ranking (launch)	eu.36kr.com ↗	2026-05-26	verified
ByteDance releases Doubao Seed Code — first coding model, SOTA on SWE-Bench	36Kr ↗	2026-05-26	verified
Doubao Seed Code integrated into Trae IDE	36Kr ↗	2026-05-26	verified
What is the difference between Doubao Seed Code and Doubao Seed 2.0 Code?	eu.36kr.com ↗	2026-05-26	to verify