DeepSeek launches V4.1-Flash and retires V4-Pro, its flagship model

DeepSeek’s V4.1-Flash is a 552-billion-parameter causal encoder-decoder that activates 8 billion parameters per input token and 16 billion per output token, supports up to a 1 million-token context, and includes native image understanding while using 890 bytes per token for KV cache. The company will retire V4-Pro on 14 September, redirecting all requests to V4.1-Flash at lower rates. Benchmark scores are 74.2 on DeepSWE v1.1, 88.1 on CyberGym, and 36.8 on Humanity’s Last Exam.