Microsoft's new MAI models

This article covers Microsoft's announcement of two new text LLMs: MAI-Thinking-1 (1T total, 35B active parameters) and MAI-Code-1-Flash (137B total, 5B active parameters). The author discusses Microsoft's claims of high performance and low cost, particularly for code generation in GitHub Copilot and VS Code. The article also delves into the training data, noting that despite claims of clean data, the models were trained on a proprietary web crawl and Common Crawl, raising licensing concerns. The author corrects initial misinterpretations about model sizes and provides technical details from the MAI-Thinking-1 paper.