deepseek / v4-pro

Overview

V4 Pro is DeepSeek’s flagship Mixture-of-Experts text model. It’s built for long-context work, with a 1,000,000-token context window. It can answer normally or run an explicit “thinking” pass before the final answer.

It returns a normal assistant reply, and it can also return a separate reasoning stream when thinking is enabled. It also supports tool calls and constrained JSON output, which helps you build agents that must follow strict formats.

What you can build

• Codebase Q&A and large PR reviews using very long prompts
• Agentic coding assistants that call tools and then finalize an answer
• Large-document analysis like policy reviews, contract issue lists, and compliance summaries
• Structured extraction that must return valid JSON for downstream systems
• Multi-turn assistants that keep context across a session
• Long-form writing with controlled style, tone, and constraints

Inputs

• A single user request as plain text. This model is text-only and does not take image, video, or audio inputs.
• An optional stable user identifier string. Use letters, digits, hyphens, and underscores only, and keep it under 512 characters. Don’t put personal data in it.
• An optional session identifier to group multiple turns into one conversation on Wiro.
• Optional system-level instructions that set behavior at the start of a session.
• A mode choice between “thinking” and “non-thinking”.
• A reasoning depth choice (low, high, or max). Higher depth produces more reasoning tokens.
• Optional randomness controls for non-thinking generations, including temperature (0 to 2) and nucleus sampling (0 to 1).
• An optional maximum response length in tokens. The total of input plus output must fit the model’s context limits.

Outputs

• A chat-style assistant message in text.
• When thinking is enabled, an additional reasoning stream separate from the final answer.
• When tool calls are enabled, a list of tool call objects that include a function name and JSON arguments.
• A finish reason that explains why generation stopped, including normal stop, length limit, tool call, content filtering, or system resource interruption.
• Token usage totals for prompt and completion. Usage can also break out reasoning tokens and cache hit or miss tokens.

Recommended settings

• For coding and math, set temperature to 0.0.
• For data cleaning and analysis, set temperature to 1.0.
• For general chat and translation, set temperature to 1.3.
• For creative writing and poetry, set temperature to 1.5.
• If you enable thinking mode, expect sampling controls like temperature and top-p to have no effect.

Limitations

• Context is large, but it is still capped at 1,000,000 tokens per request.
• Maximum output is capped at 384,000 tokens.
• In thinking mode, temperature, top-p, and repetition penalties do not affect results.
• Some repetition-penalty style controls are deprecated in the DeepSeek Chat Completions API and may be ignored.
• If you use tool calls with thinking mode, you must preserve and pass back the prior reasoning stream on later turns, or the API can reject the request.
• Messy or inconsistent inputs can degrade results. This includes poorly formatted text, mixed encodings, and copy-pasted tables with broken delimiters.

Safety & compliance

• The API can stop or redact output when content filters trigger. You may see a content-filter finish reason.
• Don’t include private or sensitive user data in the user identifier field.
• Access through the API is governed by DeepSeek’s platform terms.
• DeepSeek documents a training and data handling approach that aims to avoid sensitive personal data, and to de-identify any user input used for training workflows.