Research, written in the open.
Long-form engineering research from inside newmind AI. Models, corpora and code are released alongside each piece.
Your model can do the task. Your coding agent may not let it.
An agent's success depends not only on how capable the model is, but also on how the harness lets it use what it can do. On forty τ³ customer service tasks, three of four models looked unable to do the work inside two coding agents, yet pass the same tasks under a loop built for conversation.
Read the article →mecellem-legal-bench: Do models apply Turkish law, or just cite it?
Forty-two Turkish legal practice questions, the rubric that grades every answer point by point, and the scores of thirteen models. Not only whether the right law is cited, but whether it is applied to the facts and carried to a conclusion.
Read the article →Training GLiNER2.5 for the Turkish KVKK Law PII task
Three open models for the KVKK personal-data task in Turkish, the synthetic corpus they learned from, and the code that produced both. Every value, its category and the person it belongs to, from one pass over a document.
Read the article →