Independent Operator & Newsletter Analysis
Microsoft’s new MAI models (Simonwillison.Net)
Summary: Microsoft announced two new text LLMs: MAI-Thinking-1 (1T total, 35B active parameters) and MAI-Code-1-Flash (137B total, 5B active). The author initially misread the parameter counts, then corrected, noting the models are trained on proprietary web crawls and Common Crawl, not on cleanly licensed data as initially suggested. The author regrets the error and highlights that despite Microsoft’s claims, the training data still relies on the same problematic web scraping as other major models.
Why it matters: This matters because it reveals the persistent gap between marketing claims about ‘clean’ training data and the reality of how frontier models are actually built, which has direct implications for copyright litigation, licensing costs, and the feasibility of independent model development.
Context: Microsoft’s announcement emphasized enterprise-grade, commercially licensed data, but the technical paper shows the models were trained on a proprietary crawl of 794 billion pages plus 24.2 billion pages from Common Crawl, with only standard filtering applied.
"The majority of our web HTML corpus comes from a proprietary crawl. After initial page discovery and selection, approximately 1.2 trillion pages are crawled and parsed. […] In addition to Microsoft standard policy Sec. 2.4, we apply UT1 block list (Prigent, 2026) to remove adult content and piracy-related domains." — SIMONWILLISON.NET
Commentary: The author’s self-correction is the real story here: even a careful observer can be misled by press releases. The MAI models’ low active parameter counts are technically impressive, but the training data pipeline remains indistinguishable from competitors’ in its reliance on unlicensed web content. For independent operators evaluating model provenance, this means the ‘clean data’ claim is effectively marketing spin until Microsoft publishes far more granular sourcing documentation.
Date: June 02, 2026 06:21 PM ET
URL: https://simonwillison.net/2026/Jun/2/microsofts-new-models/
AI Sentiment Score: Neutral (50%)
AI Credibility Score: 10.0/10 — High
Scores and text generated by AI analysis of the source article indicated.Post ID: 4f419758