Microsoft offers a pair of artificial intelligence models at home

Microsoft expands the imprint of artificial intelligence with Among two new models whose teams were completely trained at home. Mai-Voice-1 is the first model for natural speech generation in Tech Major, while Mai-1-PREVIEW depends on the text, and it is the first model for the trained foundation at the end. Mai-Voice-1 is currently used in Copilot Daily and PodCast features. Microsoft has provided Mai-1-PREVIEW for general tests on LMARNA, and will start examining in specific cases in the coming weeks.

In an interview with The leader of the Microsoft Ai Mustafa Sucalyman Department said that the models pair was developed with a focus on efficiency and cost effectiveness. Mai-Voice-1 works on a single graphics processing unit and Mai-1-PREVIEW has been trained on about 15,000 NVIDIA H-100. For context, other models, such as Xai’s Grok, lasted more than 100,000 of these chips for training. Suleiman said: “Increasingly, the art and letter of training models is to choose the ideal data and not to waste any of your fluctuations on unnecessary symbols that have not learned your model much.”

Although it is used to test internal models, Microsoft Copilot is mainly based on GPT Tech from Openai. The decision to build its own models, despite its sinking At the latest artificial intelligence company, it indicates that Microsoft wants to be an independent competitor in this field. Although this may take time to reach equal with companies that have emerged as a scent in developing artificial intelligence, Suleiman told Semafor that Microsoft has a “five -year road map that we invest in a quarter after a quarter.” With some concerns that artificial intelligence can face bubbles, the Microsoft schedule should be aggressive to ensure that walking in the independent path is useful.