Microsoft's AI group described MAI-Base-2 on 15 October, a new base language model, meaning it has been pre-trained but not yet tuned for use. Microsoft claims it beats DeepSeek-V4-Pro-Base, the strongest open-source base model, on internal science, coding and maths tests. According to reports of an internal all-hands the same day, Mustafa Suleyman went further and said it is a stronger pre-trained base than the one behind OpenAI's GPT-5.6 Sol, using a quarter of the compute. Born's, a German outlet, relied on reports of that meeting rather than a Microsoft statement, and Microsoft and OpenAI declined official comment on the remarks.
The design is a mixture of experts with about 80 billion parameters active for each prompt. The total is disputed: SiliconANGLE gives 2 trillion, while Born's Tech and Windows World reports 2.8 trillion; Suleyman's own post rounds it to 3 trillion. Training ran for 38 days on 16,000 Nvidia GB200 chips in August and September, on 50 trillion tokens, about a fifth of which went to a built-in vision module meant to improve chart reading. Unlike MAI-Base-1, which mixed expert and dense layers and had around a trillion parameters, every layer here is a mixture of experts.
Microsoft's efficiency claim, which it puts at up to 12 times the hardware efficiency of DeepSeek's model, is that pre-training needed 4 to 12 times less compute than DeepSeek's model, and it says the data was organic with no distillation from other models. The reported figures include 21% model-FLOPs utilisation, 89% goodput and a knowledge cutoff of March 2026. All of it comes from Microsoft's own evaluations, and Born's notes that detailed results are unpublished and nothing has been verified independently.
On the training recipe, Microsoft says it used gradient clipping and the open-source Muon optimiser. Markets reacted on the day: CoreWeave fell 12.83%, Nebius 12.03% and Microsoft 2.44%. The release belongs to Microsoft's push to rely less on OpenAI, in which it holds a stake valued around $135 billion, and more than a dozen in-house MAI models have shipped since last quarter.
About MAI-Base-1
MAI-Base-1 was Microsoft AI's earlier base model, released in June 2026. Microsoft calls it a strong model for its size, though smaller than the open-weight and proprietary frontier models, and it was not pre-trained natively with vision, which MAI-Base-2 was.
In June 2026 Microsoft also announced MAI-Thinking-1, a general-purpose model post-trained on top of MAI-Base-1 that it says reasons strongly. Microsoft puts MAI-Base-2's efficiency gain over MAI-Base-1 at 6.3 to 12.7 times in compute.
Jun 2026
Released
Access is the other open question. SiliconANGLE says the weights will be licensed to trusted partners through Azure, while Born's reports no public release, with rollout into Microsoft products and Foundry customers. The model is still in post-training, Microsoft will use it for a new series of language models with details soon, and MAI-Base-3 is already in preparation.