Microsoft MAI-Base-2 — 2.8T-parameter base model, 4-12x FLOPs efficiency over DeepSeek
Microsoft finished pre-training MAI-Base-2, a ~2.8T total / 80B active MoE base model; the company says it is 4-12x more FLOPs-efficient than DeepSeek-V4-Pro-Base on held-out evaluations and plans to license the weights to trusted partners through Foundry.
Why it matters
Trained in 38 days on 16,000 GB200s over 50T tokens with a public efficiency blog post; Suleyman claims it beats the GPT-5.6 base on a quarter of OpenAI's compute — Microsoft's clearest break from OpenAI dependence.
What you could build with it
An enterprise AI team could license MAI-Base-2 through Foundry as a foundation for a private, domain-specialized assistant — post-training the base on internal documentation so company knowledge never leaves the tenant.
Does it hold up?
Too early to judge — the model is still in post-training, licensed only to trusted partners via Foundry, with no public API, weights, or independent benchmarks yet.
Built with Microsoft MAI-Base-2
- Microsoft Builds 2.8T-Parameter Model to Rival Open Weightsfourweekmba · Coverage of the Oct 15 announcement plus Suleyman's town-hall claim that the base beats the model behind GPT-5.6 Sol on a quarter of OpenAI's compute, and the no-distillation licensing angle.
- MAI-Base-2: What Microsoft's 2.8T-Parameter Model Meansvandatateam · FAQ-style breakdown: the 16,000 GB200s / 38-day training run, bits-per-byte vs open-weight rivals, why there are no independent benchmarks yet, and the Foundry licensing plan.
- Microsoft reveals cutting-edge MAI-Base-2 model with 2T parameterssiliconangle · Enterprise-tech press coverage: explains the all-MoE-layers design, 16k GB200 / 38-day / 50T-token training run, and Microsoft's held-out-eval claims vs DeepSeek-V4-Pro-Base.
- Microsoft Claims New MAI-Base-2 AI Model Beats OpenAI's GPT-5.6 on Less Computezubiqo · The Suleyman angle: Microsoft AI CEO's town-hall claim that MAI-Base-2 beat the base model behind GPT-5.6 Sol ('Spud') on a quarter of OpenAI's compute — framing the growing Microsoft/OpenAI friction.
- Microsoft Built Its Own Giant AI Model, So It Leans Less on OpenAIallfacts.ai · Plain-language explainer of why the in-house 2.8T-parameter training run matters: Microsoft reducing its dependence on OpenAI's models for Azure and Copilot.
Learn more
First spotted on article: source.
More AI models
OpenAI publishes 722 AI-generated math manuscripts from an unreleased modelOpenAI released 722 mathematical manuscripts in 372 result families to GitHub under Apache-2.0, produced by an…models · JEV 0.74Tencent Hunyuan Hy3: open 295B MoE with 21B active under Apache 2.0Tencent open-sourced Hy3, a 295B-parameter mixture-of-experts model with only 21B parameters active per token, under…models · JEV 0.73NVIDIA Isaac Lab 3.0 GA + Isaac OS open runtime + Agile ONE SN humanoidMajor Isaac platform expansion at GTC Berlin Oct 21: Isaac Lab 3.0 is generally available as the open-source robot…models · JEV 0.73StepFun Step-5-Preview — open BF16 weights releasedStepFun published the full BF16 weights for its flagship Step-5-Preview agentic model (600B-parameter sparse MoE, 27B…models · JEV 0.73NVIDIA Nemotron 3.5 Retriever collection: Embed 8B/1B, Rerank 8B, Parse 631MNVIDIA's Oct 21 collection drop for agentic retrieval: Nemotron-3.5-Embed in 8B BF16, 1B BF16 and 1B NVFP4…models · JEV 0.72Kimi K3.1 Pro: Moonshot's 2.8T open-weight flagship officially launchesMoonshot AI officially launched Kimi K3.1 Pro on Oct 23, a mixture-of-experts model with 2.8 trillion total / 104…models · JEV 0.72
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.