Simon Willison announces the newest release in DeepSeek’s V4 family, the DeepSeek V4 Flash 0731. With 304 billion parameters (167 GB on Hugging Face), it promises agentic capabilities that punch well above its weight. Artificial Analysis ranked it ahead of Mini Max M3, a 428B model, indicating superior performance.

Pricing is compelling: $0.14/million input and $0.27/million output. This makes it potentially the best value‑per‑intelligence model available. On the Intelligence Index vs. Cost per Intelligence Index Task chart, the model shines, especially when the reasoning level is set to high. Using Open Router, a default reasoning level yielded a disappointing pelican, but boosting reasoning effort produced a much better result, demonstrating the importance of configuration.

The release underscores DeepSeek’s commitment to balancing scale with cost‑efficiency, positioning the V4 Flash as a viable option for developers seeking high performance without prohibitive expense.