DeepSeek Claims V4 Flash Matches Sonnet 5 and Grok 4.5 on DeepSWE
DeepSeek has claimed that its newly released DeepSeek-V4-Flash model performs on par with next-generation models like Sonnet 5 and Grok 4.5 on the DeepSWE software engineering benchmark. However, these self-reported results have not yet been officially verified by the creators of the DeepSWE benchmark. If verified, this indicates that highly efficient, lower-cost Mixture-of-Experts (MoE) models can match the software engineering capabilities of much larger, next-generation frontier models. This could significantly lower the cost of deploying advanced AI coding agents. DeepSeek-V4-Flash is an efficiency-focused MoE model featuring 284 billion total parameters with 13 billion activated parameters, supporting a 1-million-token context window. The DeepSWE benchmark evaluates coding agents on long-horizon, real-world software engineering tasks while minimizing data contamination.
## BACKGROUND
DeepSWE is a contamination-free, long-horizon benchmark designed to test the coding capabilities of frontier AI agents on complex, multi-step software engineering tasks. DeepSeek is an AI research company known for releasing highly cost-effective open-weights models, and V4-Flash is part of their latest model series optimized for fast inference.