DeepSeek V4 Flash 0731

DeepSeek V4 Flash 0731 emerges as a significant advancement in open-weight foundation models, superseding its preview version with substantially enhanced agentic capabilities. This Mixture-of-Experts model boasts 284 billion total parameters, with 13 billion activated per token, and an impressive 1-million-token context window. It notably outperforms the DeepSeek-V4-Pro (Preview) on various benchmarks, despite having a far smaller activated parameter count, and is competitive with leading proprietary models. The 0731 release focuses on post-training optimizations, leading to considerable gains in coding, tool-use, and automation benchmarks. For example, on the DeepSWE benchmark for resolving real repository issues, an independent evaluator confirmed improvements, with the model also demonstrating a 12% reduction in output tokens needed for comprehensive evaluations compared to its predecessor.