DeepSeek V4

DeepSeek V4 is DeepSeek's fourth-generation model line. DeepSeek released and open-sourced the V4 preview on April 24, 2026 in two mixture-of-experts sizes: V4-Pro, with 1.6 trillion total and 49 billion active parameters, and V4-Flash, with 284 billion total and 13 billion active, both with a 1M-token context and thinking and non-thinking modes; 1M context became the default across DeepSeek's services.¹² OpenRouter lists a re-post-trained V4 Flash revision released July 31, 2026, and V4-Pro reached general availability on August 13, 2026.³² On September 10, 2026 DeepSeek released V4.1-Flash, which it calls the smallest model in a new architecture family: a 552-billion-parameter mixture of experts with a Causal Encoder–Decoder design, 8 billion active parameters for input and 16 billion for output, native visual understanding, and a KV cache that needs a quarter of the previous generation's HBM.⁴ The API's deepseek-flash name now serves V4.1-Flash and V4-Flash is retired, while DeepSeek says it will keep serving V4 Pro after September 14, 2026.⁵⁶ The weights are on Hugging Face under the MIT License.⁷

On OpenRouter, V4.1 Flash was the most-used model in the seven days to October 12, 2026, at 40.53 trillion tokens, with V4 Flash 0731 ninth (4.66 trillion) and the April V4 Flash twelfth (3.10 trillion), as ranked in Most-Used AI Models on OpenRouter by Weekly Tokens.⁸ DeepSeek's API lists deepseek-flash at $1.20 per 1M output tokens and deepseek-v4-pro, served as DeepSeek-V4-Pro-0813, at $3.96 at peak hours, sixth and tenth in Cheapest LLM APIs per 1M Output Tokens; off-peak rates are half.⁵ The DeepSeek-V4.1-Flash repo logged 1,272,612 Hugging Face downloads in the 30 days to October 13, 2026, as noted in Chinese Open-Weight Models: Downloads, Derivatives and Token Share.⁷