Microsoft unveils MAI-Base-2, a MoE model with 2.8T total and 80B active parameters, claiming 4x-12x more pretraining FLOPs-efficiency than DeepSeek-V4-Pro-Base (Microsoft AI)
In June, we announced MAI-Thinking-1, a general-purpose foundation model post-trained on top of MAI-Base-1 with strong reasoning performance.
On June 1st, Microsoft unveiled MAI-Base-2, a large language model boasting an astounding 2.8 trillion total parameters and 80 billion active parameters. This model claims to be 4-12 times more efficient in terms of pre-training FLOPs compared to DeepSeek-V4-Pro-Base.
MAI-Base-2 is an MoE foundation model, having been trained on top of MAI-Base-1, Microsoft's previous state-of-the-art general-purpose foundation model. The new model outperforms every open-weight model in Microsoft's testing across various held-out evaluations, with a lower bits-per-byte (BPB) score for text and code. For STEM, math, and Q&A tasks, evaluations were commissioned from vendors to ensure data freshness.
What sets MAI-Base-2 apart is Microsoft's full control over the model pipeline, from data provenance to training compute. This allows the company to offer a unique, trusted model to its customers via the Foundry platform, which provides enterprises with access to high-quality intelligence.
The model's impressive efficiency is due to a combination of factors, including the use of 16k GB200 GPUs for 38 days of training and a sparse, well-chosen dataset. Through a scaling ladder approach, Microsoft was able to accurately predict MAI-Base-2's performance and demonstrate a 4.2-12.3x improvement in FLOPs-efficiency over DeepSeek-V4-Pro-Base.
Written by urgent.news from Techmeme's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
