LLM & ChatActive
deepseek / v4-flash
v4-flash
DeepSeek V4 Flash is a fast MoE text model with a 1M-token context window and optional thinking mode. It’s built for chat, coding, and agent workflows.
Partner LLMDeepseekLLM Tool Call
Model ID
v4-flash
Provider
deepseek
Updated
1788534387
wiro playground—deepseek/v4-flash
Updated 1788534387
Overview
DeepSeek V4 Flash is a Mixture-of-Experts text model from DeepSeek. It supports up to 1,000,000 tokens of context and can run in fast mode or a deeper “thinking” mode. It produces assistant messages for chat and can also return strict JSON when you request structured output. This helps you turn messy text into reliable fields for downstream code and automation.
What you can build
- Long-document Q&A over large transcripts, specs, or logs
- Coding help like refactors, bug triage, and code review notes
- Agent loops that call tools, read tool results, and plan next steps
- Structured extraction to JSON for forms, tickets, and knowledge bases
- Multi-turn chat apps with persistent session history
Inputs
- A required text request describing what you want, written in plain language.
- Optional conversation identifiers so the app can keep multi-turn history consistent across requests.
- Optional “system-style” instructions that set behavior and formatting rules for a session.
- An optional reasoning toggle to choose fast replies or deeper reasoning before the final answer.
- An optional reasoning effort level that controls how much internal reasoning the model performs.
- Optional creativity controls that change how deterministic or varied the text is.
- Optional repetition and novelty controls to reduce loops and encourage new topics.
- An optional maximum response length in tokens, with a very large upper limit.
Outputs
- A primary assistant answer as text.
- When thinking is enabled, a separate reasoning channel may be returned alongside the final answer.
- When you request strict structured output, the assistant returns a valid JSON object as text.
- For tool-based agent flows, the model can emit tool call intents plus a final user-facing message.
Recommended settings
- For local open-weights style sampling, DeepSeek’s guidance is a creativity value of 1.0 with a nucleus cutoff of 1.0.
- For extended “Think Max” style reasoning, plan for a very large context budget. DeepSeek recommends at least 384K tokens available for best results.
Limitations
- This model is text-only. It does not accept image, audio, or video inputs.
- If you need image understanding, DeepSeek provides a separate experimental vision variant. This listing is for the text model.
- In thinking mode, common sampling knobs may have no effect. Expect behavior to follow the thinking policy.
- Tool-using conversations in thinking mode can require careful replay of prior reasoning messages. Missing them can break multi-turn tool flows.
- Very long inputs can still lose details, especially with repetitive sections or weak structure.
- Low-quality or messy text (OCR errors, broken tables, mixed encodings, or copied UI text) can cause extraction mistakes.
Safety & compliance
- Use of the hosted API is governed by DeepSeek’s platform terms.
- Don’t submit secrets, credentials, or sensitive personal data you don’t want processed.
- DeepSeek’s V4 model card states training data comes from public sources and licensed datasets, with efforts to filter personal and sensitive data.
- The model card also notes some user input may be used in optimization workflows, with encryption and de-identification safeguards.
API quick start
Run v4-flash with a single API call.
POST https://api.wiro.ai/v1/Run/deepseek/v4-flash
{
"prompt": "Tell me about Türkiye",
"user_id": "...",
"session_id": "...",
"systemInstructions": "..."
}