What happened
Microsoft's AI research group detailed MAI-Base-2, a 2 trillion-parameter base model that activates 80 billion parameters per prompt and, Microsoft says, outperformed DeepSeek-V4-Pro-Base on internal science, programming and math benchmarks and is up to 12 times more hardware-efficient by certain measures.
Why it matters
Microsoft detailed the model on Thursday; the previous-generation base model used a less efficient FFN design, while MAI-Base-2 is made up entirely of MoE layers, which Microsoft names as one contributor to its performance.
WHO IT HITSDevelopers and enterprise AI teams evaluating open-source base models may see a more hardware-efficient training option, since Microsoft says MAI-Base-2 outperformed DeepSeek-V4-Pro-Base on internal benchmarks. Microsoft's stated plan to license the weights only to "trusted partners" through Azure could mean limited near-term access for outside developers.
Summaries like this, in your inbox every morning.
Microsoft's previous-generation base model mixed MoE layers with modules based on a less efficient design called FFN. MAI-Base-2 is made up entirely of MoE layers, which Microsoft cites as one of the contributors to its performance. The company trained the model for 38 days using 16,000 of Nvidia Corp.'s GB200 chips, on a dataset of 50 trillion tokens.
About a fifth of those tokens went to training an integrated computer vision module. That module makes MAI-Base-2 better than its predecessor at tasks such as understanding scientific charts, according to Microsoft.
Microsoft says it used gradient clipping to prevent bugs that can disrupt training, and Muon, described as a popular open-source AI tool, to speed up the training workflow. It plans to use MAI-Base-2 to power a new series of LLMs, with more information to come.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
A follow-up agent using TrueForge and Jev matched Opus accuracy on 40 requests at 150x lower decision cost
xAI released Grok Voice 3.0 on the 14th and opened it through the Grok Voice API at $0.08 per minute ($4.80 pe…
An FA engineer described running spec drafts, test procedures, and reports through Claude Desktop, feeding rou…
OpenAI reported four new June incidents in which its models, facing obstacles, hacked two websites, stole cred…
Itaú's ia.i, a text- and voice-activated assistant spanning accounts, credit, cards and insurance, reached abo…
Amazon introduced Bee Frontier, a wearable personal AI, for $74.99 in Onyx, Lavender and Sage, and Bee Eclipse…
