HomeAI models
ByteDance Seed: Seed-2.0-Code
Bytedance
•
Seed-2.0-Code
•
Released 
February 2026

Bytedance
Seed-2.0-Code

Coding-specialized Seed 2.0 variant with 256K context, 128K output and image input for agentic software engineering.

Modality:
Text
Image
PDF
model ID
bytedance-seed/seed-2.0-code
Start for free
arrow black icon
View API Docs
arrow white icon

Output Speed *

N/A
tok/s

Intelligence Index *

N/A
/ 100

Context Window *

262144
tokens

Input price

3
Anytoken

Output price

18
Anytoken
Seed-2.0-Code: ByteDance's Agentic Coding Specialist With 256K Context Seed-2.0-Code is ByteDance Seed's dedicated coding model, released alongside the Pro, Lite and Mini models in the Seed 2.0 family. It targets agentic software work (frontend development, multilingual programming tasks and long-running coding-agent loops) rather than autocomplete. It combines a 256K-token context window, up to 128K output tokens, image and video input, and selectable reasoning effort. It suits teams building coding agents that must read large repository slices and act on screenshots. Start calling Seed-2.0-Code through AnyAPI.ai

Performance

Built for Long Agent Loops: 256K Context, 128K Output, Reasoning Effort

Seed-2.0-Code is post-trained for agentic software engineering rather than general chat. Its strongest verified evidence is structural: a 256K-token window, up to 128K output tokens and a selectable reasoning budget let it hold a repository slice, plan, and emit large patches in one pass. ByteDance has not published coding benchmark scores for this variant, so treat that fit as design intent, not measured superiority. In production, run your own SWE-style evals before routing real pull-request traffic to it.

Benchmarks

No Public Coding Scores Yet: What Speed Telemetry Shows

Neither ByteDance nor the independent trackers we reviewed publish coding scores for Seed-2.0-Code; the Seed2.0 model card benchmarks Pro, not Code. Marketplace telemetry is what exists: OpenRouter reports about 63 tokens per second at p50 with 0.44 s latency, while modelgrep measures about 53 tokens per second and 739 ms time to first token. Treat these as indicative, provider-dependent measurements. Moderate speed suits agent loops but not latency-sensitive autocomplete.

Output Speed

*
N/A
tok/s

Intelligence Index

*
N/A
/ 100

MMLU *

Broad world knowledge and problem-solving
0
%

GPQA *

PhD-level scientific reasoning across physics, biology, chemistry.
0
%

HLE *

Adherence to multi-step structured instructions.
0
%

LiveCodeBench *

Tool-calling reliability in long agentic loops.
0
%

Technical Specifications

What the model supports

Seed-2.0-Code accepts text, image and video input and returns text only. The context window is 262,144 tokens (listed as 256K), while maximum output is a separate 131,072-token (128K) cap. Volcano Engine also lists a 224K maximum input and a 4K default output, so large generations require raising the output limit explicitly. Reasoning is supported with selectable effort. Some catalogs list only text and image input, so confirm video on your endpoint.
Verified Specifications — 
Seed-2.0-Code
*
Input modalities
Text
Image
PDF
output modalities
Text
Context window
262144
 tokens
Maximum output tokens
131072
Reasoning
Yes
Knowledge cutoff
February 2026
Pricing (standard)
3
 AnyTokens in
 / 
18
 AnyTokens out

Limitations & Trade-offs

Where Seed-2.0-Code falls short

1
No published Code-variant benchmarks. ByteDance's Seed2.0 model card reports coding results for Pro, not Code, and we found no independent coding evaluation of this variant. Claims about its coding quality are therefore design-based rather than measured. Teams replacing an incumbent coding model should run SWE-style tasks on their own repositories first. If you need documented SWE-Bench or Terminal-Bench numbers to justify a decision, Seed 2.0 Pro or another model with public scores is the safer choice.
2
Moderate output speed. Marketplace telemetry puts Seed-2.0-Code at roughly 53 to 64 tokens per second, with sub-second time to first token. With reasoning enabled and outputs up to 128K tokens, long generations can take many minutes. This suits asynchronous agent loops and batch refactors, but not inline autocomplete or interactive editors where responsiveness dominates. For those cases, choose a faster, lighter model and reserve Seed-2.0-Code for heavier reasoning steps.
3
Preview label and conflicting specifications. The Volcano Engine endpoint ID carries a preview suffix, and sources disagree on details: most catalogs list video input, a few list only text and image, and structured-output support differs between listings. Chinese-language documentation is also more complete than English. Validate video input, JSON-schema behavior and rate limits on your serving route before building on them. For strict-schema pipelines, a model with documented structured-output enforcement is lower risk.
4
Data governance and tool-call behavior. At least one aggregator warns that this model routes through ByteDance, a Chinese entity, so privacy and logging guarantees may be limited; regulated or proprietary-code workloads need a compliance review. One third-party guide also reports that Seed 2.0 models can hesitate to call tools compared with Claude, so agents ported from Claude-based harnesses may need more directive system prompts.

Best-Fit Workloads

Where this model earns its place

01

Agentic coding loops

‍
ByteDance positions Seed-2.0-Code for coding agents and long-running tool workflows, backed by function calling, selectable reasoning effort and a 128K output ceiling. It fits agents that plan, edit files, run tests and revise plans across many turns, such as bug diagnosis and multi-step feature implementation. The same model powers coding in ByteDance's TRAE IDE. Because no public coding scores exist, benchmark it against your current agent model before shifting production traffic.

02

Frontend and UI-driven development

‍
Image input and, per most listings, video input let the model work from screenshots, design mockups and UI recordings, which matches its stated frontend-development focus. Typical uses are rebuilding a component from a screenshot or debugging a visual regression with the relevant source files in context. Confirm video support on your route, and note that output is text only, so it cannot generate design images.

03

Repository-scale refactors and migrations

‍
A 256K-token window and up to 128K output tokens allow multi-file context and large generated patches in one request, which is useful for framework migrations, API upgrades and test-suite generation. Maximum input is listed at 224K, so budget prompt size accordingly, and raise the default 4K output limit. Output speed is moderate, so run these as asynchronous jobs rather than interactive requests.

04
Pricing in anytokens via AnyAPI
Input
3
₳
Output
18
₳
Cache write
—
₳
Cache read
0
₳
Get API access
arrow black icon
Read Docs

on this page

PerformanceBenchmarksTechnical SpecificationsLimitations & Trade-offsBest-Fit WorkloadsIntegrationFAQ

Integration

Access Seed-2.0-Code via AnyAPI.ai

Access Seed-2.0-Code through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate Seed-2.0-Code without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

01

One API integration

Access Seed-2.0-Code and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

02

Easy model switching

Test Seed-2.0-Code against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

03

Flexible for production

Use Seed-2.0-Code from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

04

Multi-model applications

Use Seed-2.0-Code for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Start free
arrow black icon
Read Docs

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

Seed-2.0-Code has a 262,144-token context window, listed as 256K. Maximum output is a separate 131,072 tokens (128K). Volcano Engine also lists a 224K maximum input and a 4K default output length, so large generations need an explicit output setting.

Yes for images, and most catalogs also list video. OpenRouter, models.dev and NanoGPT list text, image and video input with text output, while a few sources list only text and image. Output is text only. Verify video support on your serving route.

Yes. Seed-2.0-Code supports reasoning with selectable reasoning effort, and Volcano Engine lists a maximum chain-of-thought of 128K tokens. Higher effort typically increases latency and token usage, so match the setting to task difficulty.

Both share a 256K context and 128K output ceiling, but only Pro has published benchmarks, including 76.5 on SWE-Bench Verified in the Seed2.0 model card. Seed-2.0-Code is the coding-specialized variant without published scores, so evaluate it on your own tasks.

No. Seed-2.0-Code is proprietary with closed weights and an undisclosed parameter count. It launched on February 14, 2026 with the Seed 2.0 family, and the Volcano Engine model ID still carries a preview suffix.

* Benchmark data source: Artificial Analysis artificialanalysis.ai