Home / Models / DeepSeek-V4-Pro

DeepSeek-V4-Pro

DeepSeek · LLM model

DeepSeek-V4-Pro is DeepSeek's largest model, a 1.6-trillion-parameter mixture-of-experts system with a 1M-token context and both thinking and non-thinking modes. An August 13, 2026 update (V4-Pro-0813) made it generally available in the app, web and API with stronger agent skills.

Current Open weights
ReleasedApr 24, 2026
Context window1M tokens
Input → outputtext → text
API pricedeepseek-v4-pro per 1M tokens: input cache hit $0.022 off-peak / $0.044 peak; input cache miss $0.66 off-peak / $1.32 peak; output $1.98 off-peak / $3.96 peak
Consumer accessAvailable in the DeepSeek app and on the web (chat.deepseek.com), free to use

Best for

Hard reasoning, coding and agentic tasks at a fraction of US frontier-model prices.

Overview

DeepSeek-V4-Pro is DeepSeek's largest model, a 1.6-trillion-parameter mixture-of-experts system with a 1M-token context and both thinking and non-thinking modes. An August 13, 2026 update (V4-Pro-0813) made it generally available in the app, web and API with stronger agent skills.

Strengths and limits

  • 1.6T total / 49B active MoE with 1M-token context
  • Strong coding scores (SWE Verified 80.6 on the preview card)
  • MIT-licensed weights; OpenAI- and Anthropic-compatible API
  • Text only (no image input)
  • Current API version (0813) appears to be API/app-only; published weights are the April preview

Published benchmark scores

SWE Verified (Resolved)80.6 Source: huggingface.co
GPQA Diamond (Pass@1)90.1 Source: huggingface.co
Terminal Bench 2.187.9 Source: api-docs.deepseek.com

Scores are as published by the source; settings (reasoning effort, tools, attempts) differ between labs, so compare with care. How to read benchmarks →

Sources

Facts checked Oct 14, 2026
  1. api-docs.deepseek.com/quick_start/pricing
  2. api-docs.deepseek.com/updates
  3. huggingface.co/deepseek-ai/DeepSeek-V4-Pro
  4. api-docs.deepseek.com/news/news260910