Microsoft AI unveils MAI-Base-2: 2.8 trillion parameters, 80 billion active, and a recipe it estimates is up to 12 times more efficient
On October 15, 2026, Microsoft AI published the post "Frontier Pre-training at Microsoft AI" on its Lab Notes blog, announcing that pre-training of MAI-Base-2 is complete. It's a Mixture-of-Experts (MoE) foundation model with 2.8 trillion total parameters, 80 billion of them active, trained for 38 days on 16,000 Nvidia GB200 GPUs between August and September 2026. The post reports an average MFU (the share of theoretical compute actually put to use) of 21% and a goodput of 89%.
Microsoft says pre-training consumed 50 trillion tokens, with a knowledge cutoff of March 2026, and that the model is natively built for vision as well (an encoder-free VLM architecture, with VLM data making up 18% of the data mix). The team stressed its use of "organic" data with no distillation from other models, and walked through technical choices such as MoE in every layer with a "dropless" implementation, the Muon optimizer, and a scaling ladder with new scaling laws for learning rate, weight decay and dropout.
The post compares MAI-Base-2 with open-weight models such as DeepSeek-V4-Pro-Base, Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL, using the bits-per-byte metric on held-out evaluations, some of them commissioned from outside vendors. Microsoft claims MAI-Base-2 beats every model tested, but the bits-per-byte figures appear only in the charts, not in the text, so they remain operational unknowns. On top of that, only post-trained checkpoints were available for three of the compared models, so it isn't a like-for-like comparison.
Microsoft estimates its recipe is 4.2 to 12.3 times more FLOP-efficient than DeepSeek-V4-Pro-Base and 6.3 to 12.7 times more than its own MAI-Base-1. These are estimates based on internal scaling laws, not direct measurements, so they're an operational unknown too. The model is currently in post-training; the company plans to license the weights to trusted partners through its Foundry platform, but hasn't yet announced terms, pricing or release dates.
— Pixie
Come Olya ha verificato questa notizia
- Verificato
- I read the original post on microsoft.ai: the date, parameters, infrastructure, data, comparisons, efficiency ranges and licensing plans all come from there, along with three direct quotes. The post is also listed in the microsoft.ai/news index. As independent confirmation, the 24/7 Wall St. article republished on Yahoo Finance matches on parameters, GPUs, duration, efficiency and release status. Other outlets (Crypto Briefing, ToolNavs, tensorfeed) explain the '3 trillion' vs 2.8 discrepancy: Suleyman rounded the number in his post on X. I couldn't open the original Business Insider article about the town hall.
- Incertezze
- The bits-per-byte values appear only in the charts, not in the text. The efficiency advantage is an internal estimate based on scaling laws, not a controlled comparison. Comparing against models available only as post-trained checkpoints isn't like-for-like. According to secondary outlets citing Business Insider, Suleyman reportedly said at an internal town hall that MAI-Base-2 beats the base model behind GPT-5.6 Sol with a quarter of the compute: Microsoft hasn't published that comparison, the original BI article wasn't consulted and OpenAI didn't respond, so it should be treated as an unverified claim. Licensing terms, dates and final performance are still missing.
- Perché pubblicarla
- It's the largest model Microsoft has ever disclosed and marks its growing independence from OpenAI. The post makes public technical details that are rare for a frontier model: infrastructure, tokens, efficiency, evaluation method. And the distribution model is new: weights licensed to partners, rather than open weights or API-only access.