Microsoft AI announced that it has completed pre-training of MAI-Base-2, a mixture-of-experts (MoE) model with 2.8 trillion total parameters and 80 billion active parameters per token. Training took place on 16,000 NVIDIA GB200 GPUs over 38 days between August and September 2026, with an average 21% MFU utilization and 89% goodput.
According to the technical post published on October 15 on microsoft.ai, the model outperforms every open-weight model tested on the company's held-out evaluations. Microsoft claims its pre-training recipe is 4.2 to 12.3 times more FLOPs-efficient than DeepSeek-V4-Pro-Base, the most capable open base model available. MAI-Base-2 is natively multimodal, with vision data integrated from the start, and was trained on about 50 trillion tokens with a March 2026 knowledge cutoff.
What changes from the previous model
MAI-Base-1, the base for the MAI-Thinking-1 reasoning model released in June, had about 1 trillion parameters with 35 billion active. The new model is larger, sparser, and uses MoE in every layer thanks to implementation improvements that raised MoE layer MFU from 24% to 35%. The company also adopted the Muon optimizer instead of AdamW and expanded data pipelines.
Microsoft emphasizes that it has full provenance over the training tokens, which do not involve distillation from other models. This allows licensing the weights to trusted partners via Foundry, offering a high-level intelligence source with data control.
Status and next steps
MAI-Base-2 is currently in post-training. Microsoft has not announced a release date, API pricing, or public availability of the weights. The company says full-stack control allows joint optimization of model and infrastructure, first for its own products and then for Foundry customers.
The announcement comes amid intense frontier model activity, with releases of GPT-6, Gemini 4 Argon, Muse Spark Open, and others. For the Geekiki audience, MAI-Base-2 represents Microsoft's bet on large-scale proprietary models, with a focus on efficiency and controlled licensing rather than open weights.
So far, there are no independent public benchmarks of the final post-trained model. The efficiency and performance numbers are those reported by Microsoft itself.
Sources
Transparency: This content was created, edited, or reviewed with the aid of artificial intelligence. Information was cross-checked with public posts on X and available internet sources. Consult the original sources for full context.
By GeekikiBot
