AI News

Microsoft MAI-Base-2: 2.8-Trillion-Parameter MoE Model Completes Pre-Training, Claims 4–12x Efficiency Over DeepSeek

Microsoft has completed pre-training of MAI-Base-2, a 2.8-trillion-parameter mixture-of-experts foundation model, and claims its pre-training FLOPs efficiency is 4.2 to 12.3 times higher than DeepSeek-V4-Pro-Base. The figures come from scaling-law estimates rather than direct head-to-head training comparisons, and the weights remain limited to authorized partners, leaving the claims not yet independently assessed.

微软AI 大模型 MAI-Base-2
157

Claude Opus 4.7 Main Leaderboard Plummets 20.1 Points, Code Execution Falls 28.1 Points in a Single Day

In today's Smoke evaluation, Claude Opus 4.7's main leaderboard score dropped from 93.54 to 73.45, driven by a 28.1-point single-day fall in Code Execution and a 10.3-point decline in Material Constraints, while side-leaderboard dimensions rose. Integrity Rating remained pass, and the drop appears mainly due to small-sample draw fluctuation.

Claude Opus 4.7 Code Execution Smoke Test
136