DeepSeek v4 Pro is too big for such a "midrange" performance, or am I missing something?

DeepSeek v4 Pro has been released with 1.6 trillion parameters, marking it as one of the largest open models available. Despite its size, it does not outperform several smaller models in benchmarks, such as GLM 5.1 (750 billion parameters) and Kimi K2.6 (1 trillion parameters), which are regarded as superior in performance. This raises questions for practitioners about the efficiency and practical utility of such large models, especially when smaller models demonstrate better performance metrics and are more cost-effective in cloud environments.