--- id: "2026-10-15-microsoft-mai-base-2-pretraining" url: "https://postcutoff.com/e/2026-10-15-microsoft-mai-base-2-pretraining/" as_of: "2026-10-16T12:39:00+02:00" date: "2026-10-15" date_precision: day category: model-release importance: 4 confidence: high status: [Confirmed] sources: 4 editor: human editor human_review: null version: "2026-10-16" --- As of: 2026-10-16 12:39 CEST. Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/ Canonical page: https://postcutoff.com/e/2026-10-15-microsoft-mai-base-2-pretraining/ # Microsoft AI's MAI-Base-2 base model Full title: Microsoft AI unveils MAI-Base-2, a 2.8T-parameter base model it says is 4–12x more pre-training-efficient than DeepSeek-V4-Pro-Base On Oct 15, 2026 Microsoft AI described MAI-Base-2, an in-house pre-trained mixture-of-experts base model (2.8T total, 80B active parameters, natively multimodal) trained on 50T tokens on 16k GB200 GPUs in 38 days, with no distillation from other models. Microsoft says it beats every open-weight model it tested and is an estimated 4.2–12.3x more FLOPs-efficient than DeepSeek-V4-Pro-Base. Mustafa Suleyman reportedly told staff it is a stronger base model than the one behind OpenAI's GPT-5.6, built with about a quarter of the compute. It is still in post-training and not released. ## Key facts - Architecture: mixture-of-experts, 2.8T total / 80B active parameters; encoder-free vision-language design, pre-trained natively with vision - Data: 50T-token training mixture; vision-language data 18% of the mixture (counting repetition) from 4.5T unique tokens; unique-token corpus ~36% larger than MAI-Base-1; web knowledge cutoff March 2026 - 'Our pre-training is built on organic data which does not involve distillation from other models'; crawler respects robots.txt; sources on the USTR Notorious Markets list excluded - Compute: 16k GB200 GPUs for 38 days (August–September 2026); average MFU 21%, goodput 89% - Claims (bits-per-byte on held-out text, held-out Microsoft code, vendor-commissioned STEM/maths/Q&A sets): beats every open-weight model tested; 4.2–12.3x more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base; 6.3–12.7x over MAI-Base-1. Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL were compared only as post-trained checkpoints - Methods named: scaling ladders with 'source scaling', fitted hyperparameter scaling laws, dropless MoE with fixed-size memory (MoE-layer MFU 24% → 35%), Muon optimizer with resharding, activation gradient clipping - Availability: none yet; post-training under way, weights to be licensed to 'trusted partners' through Foundry; MAI-Base-3 in preparation - Suleyman to employees (reported Oct 15–16, Business Insider via secondary coverage): stronger than the OpenAI base model behind GPT-5.6, using about a quarter of its training compute; based on Microsoft's own information about OpenAI's models; OpenAI did not comment ## What happened In a Lab Notes post on Oct 15, 2026, Microsoft AI described **MAI-Base-2**, the base (pre-trained, not yet post-trained) model behind its next generation of in-house models. It is a sparse mixture-of-experts with 2.8 trillion total and 80 billion active parameters. It was trained on a 50-trillion-token mixture that includes images from the start, on 16,000 Nvidia GB200 GPUs over 38 days in August and September. Microsoft stresses that its data is "organic" and not distilled from other models. Microsoft measures quality as bits-per-byte on held-out text, its own unpublished code and commissioned STEM and Q&A sets. On that basis it says MAI-Base-2 beats every open-weight model it tested. It puts the model at 4.2–12.3 times more FLOPs-efficient in pre-training than DeepSeek-V4-Pro-Base, the strongest open base model it could test, and 6.3–12.7 times more than its own MAI-Base-1. The post does not list per-benchmark scores in text; they are in figures. Business Insider then reported that Mustafa Suleyman told employees MAI-Base-2 is a stronger base model than the OpenAI model behind GPT-5.6, trained with about a quarter of the compute. This entry relies on secondary coverage of that report, since the Business Insider article was not read directly. The comparison rests on Microsoft's own information about OpenAI's models, and OpenAI did not comment. ## Why it matters Under the restructured Microsoft–OpenAI deal, Microsoft can build frontier models of its own. This is the first time it has given detailed pre-training numbers at frontier scale (2.8T parameters, ~16k GB200s) and said publicly that its base model beats its partner's. All efficiency figures are Microsoft's own, measured on evaluation sets it chose and has not published, so they are unverified. The model is not available yet. Microsoft says it plans to license weights to trusted partners rather than release them openly. ## Your AI and this story - GPT-6 Astra (training cutoff April 2026): 168 days after its cutoff - Claude Opus 5.5 (training cutoff June 2026): 107 days after its cutoff - Gemini 4 Argon (training cutoff March 2026): 198 days after its cutoff - Grok 4.7 (training cutoff May 2026): 137 days after its cutoff ## Sources 1. [Microsoft AI Lab Notes: Frontier pre-training at Microsoft AI (MAI-Base-2)](https://microsoft.ai/news/frontier-pre-training-at-microsoft-ai/) (microsoft.ai, official) 2. [Startup Fortune: Microsoft AI chief says his model beat OpenAI's using a quarter the compute](https://startupfortune.com/microsoft-ai-chief-says-his-model-beat-openais-using-a-quarter-the-compute/) (startupfortune.com, press) 3. [Digital Today: Microsoft AI chief says company built model stronger than GPT-5.6 with quarter of computing power](https://www.digitaltoday.co.kr/en/view/115636/microsoft-ai-chief-says-company-built-model-stronger-than-gpt-56-with-quarter-of-computing-power) (digitaltoday.co.kr, press) 4. [Techmeme thread (Oct 15)](https://www.techmeme.com/261015/p59) (techmeme.com, discussion) ## Changes - 2026-10-16 (filed): Created (Microsoft AI post read directly; Suleyman remarks from secondary coverage of Business Insider) ## Related - 2026-06-02: [Microsoft launches seven in-house MAI models at Build 2026, led by MAI-Thinking-1](https://postcutoff.com/e/2026-06-02-microsoft-mai-models-build-2026/index.md) - 2026-04-27: [Microsoft and OpenAI restructure partnership, drop the AGI clause and exclusivity](https://postcutoff.com/e/2026-04-27-microsoft-openai-deal-restructured/index.md) - People: [Mustafa Suleyman](https://postcutoff.com/person/mustafa-suleyman/)