AI model · by DeepSeek
DeepSeek V4 Pro
DeepSeek V4 Pro is DeepSeek's model for agentic tasks; the company says its performance gains are most noticeable in production settings.
- 1M tokens context
- Open weights · MIT
- Checked against DeepSeek's page 19 Oct 2026
What DeepSeek V4 Pro is for
DeepSeek says it built DeepSeek V4 Pro with much stronger agentic skills, and that its performance gains are most evident in real production use. The company directs readers to the inference providers it lists for the model and to the instructions in its inference folder for running DeepSeek-V4 locally.
What DeepSeek says it does well
- Outperforms DeepSeek-V4-Pro (Preview) on the benchmarks listed on the page.
- Includes DSpark speculative decoding module attached to the DeepSeek-V4-Pro (Preview) structure.
Limits DeepSeek mentions
- No Jinja-format chat template is included; a Python encoding folder is provided instead.
- For high and max reasoning effort levels, a maximum output length of 384K tokens is recommended.