Mirendil-Base-Large Achieves Lowest Perplexity on Private Codebases
October 23, 2026
Baseten's testing of the Mirendil-Base-Large model on unseen private codebases showed it achieved the lowest bits per byte (perplexity) of all tested models. This indicates superior efficiency in understanding complex, novel code structures.
HOW THIS AFFECTS YOU
●
builderYou should evaluate Mirendil models for code-generation tasks where perplexity and accuracy are critical.
●
researcherThis provides a benchmark for evaluating model efficiency on zero-shot private datasets.