DeepSeek-V4-Flash-0731 Outperforms DeepSeek-V4-Pro-Preview in Benchmarks
The DeepSeek-V4-Flash-0731 model has reportedly surpassed the larger DeepSeek-V4-Pro-Preview model in benchmark tests. This performance boost was achieved through re-post-training, while keeping the model's original architecture and size intact. This development demonstrates that smaller, more cost-effective models can achieve superior performance through optimized post-training techniques. It highlights a significant industry trend toward model efficiency, allowing developers to access high-level capabilities at a fraction of the cost. DeepSeek-V4-Flash-0731 maintains a Mixture-of-Experts (MoE) architecture with 284B total and 13B active parameters, and is priced competitively at $0.14 per million input tokens and $0.28 per million output tokens. Both the Flash and Pro versions support three reasoning effort modes to customize performance.
## BACKGROUND
DeepSeek-V4 is a family of large language models that utilizes a Mixture-of-Experts (MoE) architecture, where only a subset of parameters (active parameters) is activated for each token to save compute. The 'Pro' version is a larger model designed for maximum capability, whereas the 'Flash' version is a smaller, faster, and more cost-effective variant optimized for speed.