Benchmarks compare backend generation using open-source local models

A benchmark tests DeepSeek v4 Pro and other local models for Python backend code generation.