On July 31, Chinese artificial intelligence company DeepSeek announced the public beta launch of the API for its latest large language model, DeepSeek-V4-Flash. The official version maintains the same model architecture and parameter scale as the preview version, with iterative optimization during post-training, delivering a substantial leap in agent capabilities. Several code agent benchmark scores surpassed the higher-tier V4-Pro preview, including 82.7 on Terminal Bench 2.1, 54.2 on NL2Repo, and 76.7 on Cybergym. In full-stack software engineering tests, it scored 54.4 on DeepSWE and 68.7 on DSBench-FullStack. V4-Flash is a Mixture-of-Experts model with 284 billion total parameters, activating 13 billion per inference, compared to V4-Pro's 1.6 trillion total parameters and 49 billion activated. Both use a hybrid attention architecture combining Compressed Sparse Attention and Heavily Compressed Attention. Using V4-Pro as an example, in million-token long-context scenarios, compute per token drops to 27% and KV cache to 10% compared to the previous-generation V3.2. The capability leap comes entirely from reinforcement learning and distillation tuning during post-training, with no changes to architecture or parameter scale.