Evidence brief
Mustafa Suleyman says Microsoft’s first reasoning model, MAI-Thinking-1, was built from scratch and is now available on Microsoft Foundry.
Mustafa Suleyman says Microsoft’s first reasoning model, MAI-Thinking-1, was built from scratch and is now available on Microsoft Foundry.
This brief preserves the original source so the summary and editorial context can be checked independently.
Source attributionX：Mustafa Suleyman（Microsoft AI CEO） (@mustafasuleyman) · X：Mustafa Suleyman（Microsoft AI CEO） (@mustafasuleyman)
Open the original sourceRelationships are derived from shared topics, entities, categories, tags, and community context; every result remains independently source-linked.
@satyanadella reposted @mustafasuleyman’s message: MAI-Thinking-1 is their first reasoning model, built from scratch, now available in Microsoft Foundry, with thanks to the team.
Original sourceMicrosoft Research says multilingual Agent evaluations often compare only final answers and discard trajectories; the paper uses action policy as the metric, covering 8 models, 6 parallel benchmarks, 41 languages, and 2.38M rollouts, and analyzes five types of confounding factors.
Original sourceAfter finishing an RLHF textbook, the author reflected that LLM progress in long-form nonfiction writing has stalled. Strong writing models such as GPT 4.5 and Kimi K2 already feel dated, while coding, math, and similar tasks are nearing superhuman levels. Models can fix typos and edit, but they are still disorganized and error-prone when structuring entire chapters, which the author believes prevents them from autonomously solving open scientific problems.
Original source@op7418 said, based on circulating screenshots, that DeepSeek Harness will enter public beta today, and further speculated whether it will be released together with model details for 0813. The post did not provide verified external-link text.
Original sourceSGLang and Miles supported Qwen3.8-2.4T-A95B on launch day; the model is Qwen’s largest open-source model, with 2.4T total parameters, 95B active per token, and a hybrid attention architecture.
Original source