Microsoft launches MAI-Base-2, a MoE model with 2.8 T total and 80 B active parameters, claiming 4‑12× higher pretraining FLOPs‑efficiency than DeepSeek‑V4‑Pro‑Base (Microsoft AI)
deepseek microsoft training
| Source: Techmeme | Original article
Microsoft unveiled MAI-Base-2, a MoE model with 2.8 trillion total and 80 billion active parameters, claiming 4‑12× higher pre‑training FLOPs efficiency than DeepSeek‑V4‑Pro‑Base.
Microsoft AI has announced the completion of MAI‑Base‑2, its newest foundation model. The 2.8‑trillion‑parameter mixture‑of‑experts (MoE) system activates roughly 80 billion parameters per token, a design that lets it achieve the same computational work with far fewer floating‑point operations. Microsoft says the model delivers four to twelve times the pre‑training FLOPs‑efficiency of DeepSeek‑V4‑Pro‑Base, a leading competitor in the large‑model arena.
The model was trained over 38 days on a cluster of 16,000 Nvidia GB200 GPUs, ingesting 50 trillion tokens from a mixed corpus that is 18 % visual data (about 4.5 trillion image‑text pairs). By activating only a subset of its experts for each token, MAI‑Base‑2 keeps the active parameter count at 80 billion while retaining the expressive power of a 2.8‑trillion‑parameter network.
Why it matters: the efficiency claim, if borne out in downstream benchmarks, could lower the cost barrier for building and running multimodal AI systems at scale. MoE architectures promise to stretch the capabilities of foundation models without a proportional rise in compute expense, a development that could accelerate adoption across cloud services, enterprise AI, and research. For the Nordic AI community, the model’s visual‑language training mix aligns with growing interest in multimodal applications such as automated content moderation, translation, and creative tooling.
What to watch next: Microsoft has yet to detail how MAI‑Base‑2 will be made available—whether via Azure AI, open‑source release, or licensing. Performance results on standard benchmarks, especially in comparison with the earlier MAI‑Base‑1 and the June‑released MAI‑Thinking‑1 fine‑tuned model, will be closely scrutinised. Industry reactions and potential collaborations with partners in Europe will also shape how quickly the model moves from a research milestone to a production‑ready service. As we reported on 16 October, Microsoft’s “Frontier Pre‑training” effort has been steadily scaling its models; MAI‑Base‑2 marks the latest step in that trajectory.
Sources
Back to AIPULSEN