ByteDance / seed-v2.1-turbo
seed-v2.1-turbo
Seed v2.1 Turbo by ByteDance is a 256K-context reasoning model for coding and agent workflows. It can use images for understanding and returns text or JSON.
Overview
Seed v2.1 Turbo is a text generation model from ByteDance. It’s built for coding delivery and long-horizon agent tasks. You give it a written task, plus optional images for context. It produces a single assistant answer as text, and it can format outputs as strict JSON when you ask. That helps you ship reliable code changes, audits, and multi-step plans without losing context.
What you can build
- Repo-aware coding help that keeps long technical context in one thread
- Multi-step debug plans that include tests, checks, and rollback steps
- Extraction and transformation pipelines that output valid JSON objects
- Visual QA for screenshots, diagrams, and UI flows (image understanding)
- Agent-style task plans for research, office workflows, and tool-guided execution
Inputs
- A required written instruction in plain text. Include goals, constraints, and acceptance criteria.
- Optional system-level instructions that set the assistant’s role and rules for the whole run.
- Optional images for visual understanding. You can attach multiple images in one request, up to 20 files.
- Optional user and session identifiers (string or number). Provide them to keep chat history grouped correctly.
- Optional randomness control from 0 to 2. Lower values make outputs more repeatable.
- Optional nucleus sampling control from 0 to 1. Lower values reduce rare-token choices.
- Optional repetition controls from -2 to 2. Use them to discourage repeated phrases.
- Optional maximum completion length from 0 to 65,536 tokens.
- Optional reasoning depth selection: minimal, low, medium, or high.
- Optional thinking mode selection: enabled, disabled, or auto.
Outputs
The model returns a text response. The response can be plain language, code blocks, or a JSON object. When you request structured output, the model can keep the result valid JSON. For long tasks, the response may include detailed step-by-step work, plus final deliverables.
Recommended settings
- Use thinking mode disabled for short answers, classification, and strict formatting tasks.
- Use reasoning depth high for multi-file refactors, debugging, and long task plans.
- Keep randomness near 0 for deterministic JSON and repeatable code edits.
- Raise randomness toward 1 to 2 for brainstorming, naming, and creative drafts.
- Set a large completion limit only when you need long code or long reports.
Limitations
- The model outputs text only. It does not generate images, audio, or video.
- Visual inputs support understanding tasks. They don’t enable image editing or image generation.
- Very large outputs can be cut off when they hit the configured length limit.
- Low-quality inputs reduce reliability. This includes blurry screenshots, tiny text, cropped diagrams, and noisy scans.
- If your prompt mixes many goals, the model may optimize for one and miss others.
Safety & compliance
Follow Wiro’s usage rules and your local laws. Don’t submit private, confidential, or regulated personal data unless you have a clear right to use it. Don’t use the model to create malware, exploit code, or instructions for wrongdoing. If you upload images, make sure you have permission to process that content. Review outputs before you run code or act on high-stakes advice.
API quick start
Run seed-v2.1-turbo with a single API call.
{
"prompt": "Tell me about Türkiye",
"inputImage": "https://your-cdn.com/input.png",
"userId": "...",
"sessionId": "..."
}