[AINews] Reve 2 and Ideogram 4: Layouts in Imagegen
Microsoft released MAI-Thinking-1, a reasoning model trained without third-party distillation, with a 109-page transparency-heavy report.
“hillclimbed from scratch”— Mustafa Suleyman
Microsoft introduced MAI-Thinking-1, a generalist reasoning model trained with zero synthetic data and zero prior-model distillation, reporting 97% on AIME 2025, 53% on SWE-Bench Pro, and human-preference wins over Sonnet 4.6. The accompanying 109-page report drew praise for unusual transparency, disclosing the scaling-ladder recipe, MFU numbers, and data mixture. It matters as a credible frontier-model effort built from scratch without leaning on competitors' outputs.