Code-generation benchmark across three model architectures (dense LLM, Mixture-of-Experts and Small Language Model) on 5 benchmarks and 15 languages, with pass@1 and CodeBLEU metrics.
python jupyterlab mistral huggingface humaneval codebleu mbpp deepseek-v4 claude-opus-4-7 multipl-e plot2code
-
Updated
Jul 26, 2026 - Python