Seed-V2 Mini LLM
The smallest model in ByteDance's Seed 2.0 series, a reasoning LLM that reads images and calls functions. Reasoning is on by default and can be switched off for simple jobs.
Overview
Seed 2.0 Mini, listed on Wiro as Seed-V2 Mini, is the smallest of the three general-purpose models in ByteDance's Seed 2.0 series, alongside Pro and Lite. ByteDance positions it for high-concurrency and batch generation workloads.
Mini thinks before it answers unless you turn that off, and the depth of that thinking is adjustable. ByteDance lists a 256K-token context window with up to 224K tokens of input. Mini also reads images and calls functions.
In its Seed 2.0 model card, ByteDance reports 83.6 on MMLU-Pro and 79.0 on GPQA Diamond for Mini. Seed 2.0 Lite scores 87.7 and 85.1 on the same two tests.
What you can build
- Bulk text work such as ticket triage, product tagging, review summaries or dataset labelling.
- A chat assistant that remembers earlier turns when each request carries the same user ID and session ID.
- Field extraction from scanned documents and charts sent as images.
- Agents that call your functions, several in one turn when needed.
- Jobs that need JSON back, with a schema that keeps every reply parseable.
Inputs
- Prompt or messages. A single prompt or a messages array with the conversation so far, never both.
- Images (optional). Up to 20 files or URLs in JPG, PNG, GIF, WebP or HEIC.
- System instructions (optional).
- User ID and session ID (optional). Pass both to keep chat history between calls.
- Sampling: temperature 0 to 2 (default 1.0), top P 0 to 1 (default 0.7), and frequency and presence penalties from -2 to 2 (default 0).
- Max completion tokens: 0 to 65,536, default 65,536, with reasoning tokens counted.
- Reasoning effort: minimal, low, medium (default) or high.
- Thinking type: enabled (default) or disabled. ByteDance documents only these two for this model, so the auto option runs with thinking on.
- Tools: up to 128 OpenAI-style function definitions, a tool choice of auto, none, required or a named function, and parallel tool calls (on by default).
- Previous task token and tool outputs: return function results, each matched to a call ID from the earlier turn.
- Response format (optional): a JSON object, or a JSON schema with strict matching available.
Outputs
A text reply, or function calls when tools are passed. Once you return their results, the model finishes the same turn.
Recommended settings
- The default temperature 1.0 and top P 0.7 suit chat and general writing.
- For short classification or extraction, set thinking to disabled or reasoning effort to minimal. ByteDance documents both as turning reasoning off for this model.
- Use high effort for math problems and multi-step planning.
- With thinking on, keep max completion tokens generous. Reasoning draws from the same budget, so a low cap can end the reply before the answer.
Limitations
- In ByteDance's tables, Mini scores below Lite and Pro on MMLU-Pro and GPQA Diamond. ByteDance aims Pro at complex reasoning and long-context tasks.
- ByteDance also tests Mini on video benchmarks, but this page does not accept video input.
- Custom tool types are rejected. Only function tools work.
- ByteDance lists up to 128K output tokens for this model. This page caps a reply, reasoning included, at 65,536.
API quick start
Run seed-v2-mini with a single API call.
{
"prompt": "Tell me about Türkiye",
"inputImage": "https://your-cdn.com/input.png",
"user_id": "...",
"session_id": "..."
}curl -X POST "https://api.wiro.ai/v1/Run/bytedance/seed-v2-mini" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "Tell me about Türkiye",
"inputImage": "https://your-cdn.com/input.png",
"user_id": "...",
"session_id": "..."
}
JSONOther models you might like
Explore similar llm & chat models.
