DeepSeek V4 Pro
V4 Pro is a text reasoning model. The original preview has 1.6 trillion total parameters and 49 billion active parameters. Its 0813 successor retains that model structure and adds a DSpark speculative decoding module.
The general-availability release adds speculative decoding
The 0813 checkpoint superseded the April preview. Its release card provides message encoding and output parsing code instead of a Jinja chat template. Reasoning effort can be low, high or max.DeepSeek dated release historyDeepSeek V4 Pro 0813 release card
The hosted Pro endpoint accepts text. The Responses interface can run web search for Pro, but does not enable that built-in tool for Flash.DeepSeek model capability responseDeepSeek Responses API support
Max is a reasoning mode
The V4 model card uses Pro-Max for the maximum reasoning setting. It is not presented as an additional downloadable model. The architecture combines compressed sparse attention with heavily compressed attention.DeepSeek V4 Pro model card
Practical guidance. Record reasoning effort with your evaluation results. Otherwise, two measurements for V4 Pro may describe different inference budgets.
Some sampling options have no effect
In thinking mode, temperature and penalties are ignored. For Pro, top-p is fixed to 0.95 with tool calling and 1.0 without it. Earlier reasoning must be returned when the request includes tools.DeepSeek thinking-mode behavior
Practical guidance. Do not use a setting in the request log as proof that it controlled generation. Check the provider’s model-specific behavior before explaining a change in output quality.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | |
| Access | API, Downloadable |
| Model identifiers | deepseek-v4-proDeepSeek-V4-Pro-0813deepseek-ai/DeepSeek-V4-Pro |
| Context window | 1,048,576 tokens API limit. DeepSeek model capability response |
| Maximum output | 393,216 tokens API maximum output. DeepSeek model capability response |
| Input price | 0.66 USD per million tokens Off-peak, uncached input. DeepSeek models and prices |
| Input price | 1.32 USD per million tokens Peak, uncached input. DeepSeek models and prices |
| Output price | 1.98 USD per million tokens Off-peak output. DeepSeek models and prices |
| Output price | 3.96 USD per million tokens Peak output. DeepSeek models and prices |
| Input and output | Text input and text output; no native image input on this API model.DeepSeek model capability response |
| Hosted release | The deepseek-v4-pro API currently serves the 0813 release.DeepSeek models and prices |
| Published weight license | MIT.DeepSeek V4 Pro model card |
Related models
- DeepSeek V4.1 Flash
DeepSeek · DeepSeek V4.1
Multimodal mixture-of-experts model with a causal encoder-decoder design and separate computation budgets for reading input and generating text.
Launch date API · Downloadable