← All models

DeepSeek-V4-Pro-Base

Published Publisher DeepSeek

Overview

DeepSeek-V4-Pro-Base is the open-weight base language model in DeepSeek’s V4 Pro family, distinct from the post-trained DeepSeek-V4-Pro model. The V4 preview was announced on April 24, 2026; the Base repository’s history also records a V4 release commit on that date. (api-docs.deepseek.com)

Architecture and capabilities

  • Mixture-of-Experts architecture with 1.6 trillion total parameters, 49 billion activated parameters, a 1 million-token context window, and mixed FP8 precision. (huggingface.co)
  • DeepSeek reports base-model evaluations covering world knowledge, language and reasoning, code and mathematics, and long-context tasks. These are benchmark results, not guarantees of performance on individual tasks. (huggingface.co)

Uses and limitations

The base checkpoint can serve as a foundation for research, evaluation, fine-tuning, and self-hosted text-generation systems. It should not be assumed to have the instruction-following, reasoning modes, or API features of the separately post-trained Pro model. Its approximately 1.61 TB repository also makes local deployment resource-intensive. Outputs can be incorrect and should be verified for consequential use. (huggingface.co)

The repository includes an MIT License; consult the license for its conditions and disclaimer. (huggingface.co)

Pricing

As checked on November 16, 2026, a dedicated official API price for this Base checkpoint was not verified on either DeepSeek’s Chinese-language CNY pricing page or its international USD pricing page. Neither an input-token nor an output-token rate can therefore be assigned to this model. The listed deepseek-v4-pro API serves a different model version, so its prices are not applied to Base. Self-hosting may incur infrastructure costs. (api-docs.deepseek.com)

Official pricing sources: Chinese-language pricing and international pricing.

Official sources