DeepSeek V4-Flash Released

DeepSeek has launched the official version of DeepSeek-V4-Flash-0731 and transitioned its API to public beta. The core architecture remains identical to the preview, featuring 284B total parameters and 13B active parameters along with a 1 million token context window, with only post-training newly performed.

Key changes and performance metrics are as follows:

• Terminal-Bench 2.1: Increased from 61.8 to 82.7
• Artificial Analysis Intelligence Index: Increased from 40.3 to 49.9
• Price per 1 million output tokens: Maintained at $0.28
• Cached input price: $0.028 → $0.0028 per 1 million tokens
• Cost for Artificial Analysis evaluation tasks: Approximately $0.027, the lowest among compared models

The official checkpoint includes a speculative decoding module by default, allowing inference without a separate draft model. It also supports reasoning_effort settings of low, high, and max, as well as the Responses API.

The existing deepseek-chat and deepseek-reasoner model names were discontinued after July 24, 2026, and users must switch to deepseek-v4-flash or deepseek-v4-pro.