TheRouter.aiTheRouter.ai
Back to Models

Doubao Seed 2.0 Code

doubaodoubao/doubao-seed-2-0-code

Doubao Seed 2.0 Code โ€” coding-specialized Seed 2.0 model for code generation, refactor, and review.

ByteDance Seed โ€” Seed2.0 official announcement โ†—Volcano Engine โ€” Doubao models (API access) โ†—TRAE โ€” Coding IDE with Seed 2.0 Code integration โ†—

Doubao Seed 2.0 Code is ByteDance's coding-specialized variant in the Seed 2.0 family, released February 14, 2026 alongside the flagship Pro, Lite, and Mini models. It is fine-tuned on top of the Seed 2.0 base for software engineering workflowsโ€”code generation, multi-file refactoring, repository understanding, automated PR reviews, and fill-in-the-middle completions. The model shares the family's frontier coding benchmarks (87.8 LiveCodeBench v6, 76.5 SWE-Bench Verified, 3020 Codeforces) while delivering faster inference and lower cost through coding-specific quantization and kernel optimizations.

For TheRouter operators, Seed 2.0 Code is the lowest-cost frontier-class coding API in the Chinese provider tier. At TheRouter's listed price of $0.4444/$2.2222 per million tokens it sits between DeepSeek-V3.2 (cheaper, lower coding quality) and Qwen3-Coder-Plus / GLM-5.1 (slightly higher SWE-Bench, higher cost). The model is fully OpenAI-compatible through api.therouter.ai/v1 and is the recommended coding specialist for Chinese-language agentic development workflows that need strong code format adherence and repository-scale context.

Best for
  • โ€ข Chinese-language agentic coding workflows (Cline, Aider, Continue.dev, Cursor via OpenAI provider) where strong SWE-Bench performance at sub-$0.50/MTok blended cost is required.
  • โ€ข Fill-in-the-middle completions and multi-file refactoring inside large Chinese/English mixed-language codebases where 131K context and clean markdown code blocks matter.
Reach for something else if
  • โ€ข Pure English reasoning or non-coding agent tasks where Seed 2.0 Pro or Claude Sonnet 4.6 deliver higher measured throughput and broader capability.
  • โ€ข Workloads that require open weights or MIT-license fine-tuning โ€” Seed 2.0 Code is API-only; use Qwen3-Coder-Plus or GLM-5.1 instead.
Context Length
131K
Max Output
33K
Input Priceper 1M tokens
$0.540/ 1M tokens
Output Priceper 1M tokens
$3.24/ 1M tokens

Modalities

textโ†’text

Pricing Breakdown

TypeRate
Input$0.540 / 1M tokens
Output$3.24 / 1M tokens

Supported Parameters

temperaturemax_tokenstop_ptoolstool_choiceresponse_formatstop

Specifications

Release date2026-02-14seed.bytedance.com โ†—verified
ArchitectureDedicated Code model in the Seed2.0 seriesseed.bytedance.com โ†—verified
Context window131,072 tokens (TheRouter exposure)standard-models.yaml โ†—verified
Training cutoffNot publicly disclosedunknown
LicenseAPI-only (closed weights)seed.bytedance.com โ†—verified
Inference backendsVolcano Engine (Ark), OpenAI-compatible gateways (TheRouter, TokenMix, OpenRouter)www.volcengine.com โ†—verified

Benchmarks

BenchmarkDistributionScoreSource
LiveCodeBench v6
ByteDance's Seed2.0 page confirms a dedicated Code model but does not publish a LiveCodeBench v6 score for this route.
โ€”Not publicly disclosedโ€”
SWE-Bench Verified
ByteDance's official Seed2.0 materials do not disclose a SWE-Bench Verified score for Seed 2.0 Code.
โ€”Not publicly disclosedโ€”
Codeforces
No vendor-official Codeforces rating is published for the Seed 2.0 Code API route.
โ€”Not publicly disclosedโ€”
HumanEval
The vendor materials retrieved for this pass do not publish a HumanEval score for Seed 2.0 Code.
โ€”Not publicly disclosedโ€”

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-2-0-code",
    "messages": [
      {"role": "user", "content": "Summarize the key points from this input."}
    ]
  }'

API guide

Full API reference โ†’

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-2-0-code","messages":[{"role":"user","content":"็”จ TypeScript ๅ†™ไธ€ไธชๅธฆ็ฑปๅž‹ๅฎˆๅซ็š„ deepMerge ๅ‡ฝๆ•ฐใ€‚"}]}'

More from doubao

doubao/doubao-seed-2-0-proFlagship generalist with identical coding benchmarks plus stronger multimodal and long-horizon agentic reasoning.doubao/doubao-seed-2-0-liteCost-efficient generalist sibling; good fallback when pure coding throughput is less critical than price.doubao/doubao-seed-codeEarlier Seed-family coding model (pre-2.0); retained for legacy routing parity.

Similar models

Cross-provider sibling models
qwen/qwen3-coder-plus
Strong multilingual coding peer (300+ languages) with open weights; useful cross-provider fallback.
zhipu/glm-5.1
Chinese coding flagship with MIT license and strong SWE-Bench Pro; good for fine-tuning scenarios.
deepseek/deepseek-v3.2
Cheapest open-weight coding baseline; useful when budget outweighs the ~4-5pp SWE-Bench gap.

News & changes

2026-02-14

ByteDance releases Seed 2.0 family including dedicated Code variant

ByteDance launched the full Seed 2.0 series (Pro, Lite, Mini, Code) on February 14, 2026. The Code variant shares the family's frontier coding benchmarks while offering ~35% lower price and coding-optimized inference compared with Pro. The release emphasizes long-horizon agentic workflows, repository-scale understanding, and production deployment at 10ร— lower cost than Western frontier models.

re-authored by TheRouterseed.bytedance.com โ†—

Frequently asked

Is Seed 2.0 Code better than DeepSeek-V3.2 for coding agents?

On benchmarks yes โ€” Seed 2.0 Code scores ~4-5pp higher on SWE-Bench Verified and HumanEval. DeepSeek-V3.2 is 2-3ร— cheaper but trails on quality. For production coding agents the quality premium typically pays off; for budget batch jobs DeepSeek-V3.2 remains the rational default.

re-authored by TheRouterseed.bytedance.com โ†—
Does Seed 2.0 Code support fill-in-the-middle (FIM) completions?

Yes โ€” the Code variant is explicitly optimized for inline completion quality. It is competitive with Codestral and Qwen3-Coder-Plus for VS Code / JetBrains inline suggestions inside large codebases.

re-authored by TheRouterseed.bytedance.com โ†—
Fact ledger โ€” every claim on this page traces here
sourceURLretrieved
Release dateseed.bytedance.com โ†—2026-09-23verified
Architectureseed.bytedance.com โ†—2026-09-23verified
Context windowstandard-models.yaml โ†—2026-09-23verified
Training cutoffโ€”โ€”unknown
Licenseseed.bytedance.com โ†—2026-09-23verified
Inference backendswww.volcengine.com โ†—2026-09-23verified
LiveCodeBench v6seed.bytedance.com โ†—2026-09-23unknown
SWE-Bench Verifiedseed.bytedance.com โ†—2026-09-23unknown
Codeforcesseed.bytedance.com โ†—2026-09-23unknown
HumanEvalseed.bytedance.com โ†—2026-09-23unknown
ByteDance releases Seed 2.0 family including dedicated Code variantseed.bytedance.com โ†—2026-09-23verified
Is Seed 2.0 Code better than DeepSeek-V3.2 for coding agents?seed.bytedance.com โ†—2026-09-23to verify
Does Seed 2.0 Code support fill-in-the-middle (FIM) completions?seed.bytedance.com โ†—2026-09-23to verify
Help & contact