AI Tools: Microsoft Azure Adds DeepSeek; Alibaba Releases New Model; Meta Gets ‘Memories’

This week in artificial intelligence models and tools, Microsoft announced that DeepSeek R1 is available on Azure AI Foundry, and Alibaba released a model that it said is better than DeepSeek. Meanwhile, Meta launched a memory feature, and Google announced new Gemini Live capabilities.

Microsoft is making an AI model from Chinese startup DeepSeek available to its Azure cloud customers.

The software giant said in a Wednesday (Jan. 29) blog post that DeepSeek R1, the startup’s reasoning model, is joining the Azure AI Foundry, a one-stop-shop platform for designing, customizing and managing AI solutions. R1 will be among the more than 1,800 models available for use in the foundry.

DeepSeek R1 on Azure resides on a “trusted, scalable and enterprise-ready platform” backed by the tech giant’s reliability, per the post.

“R1 offers a powerful, cost-efficient model that allows more users to harness state-of-the-art AI capabilities with minimal infrastructure investment,” wrote Asha Sharma, Microsoft’s corporate vice president of AI platform, in the post.

Meanwhile, DeepSeek caused chip stocks to tumble Monday (Jan. 27) amid media reports that it only spent $5.6 million and used about 2,000 slower H800 Nvidia chips to train its V3 model. Over several days, DeepSeek shot to the top of the Apple App Store’s most downloaded apps.

However, some analysts said other costs were not included in the tally. Microsoft and OpenAI also suspected DeepSeek of using a technique called distillation, in which outputs of a larger model — from OpenAI — are used to guide the smaller model to achieve similar performance. This would violate OpenAI’s terms of service.

There were also concerns about Chinese censorship on DeepSeek, but since it is open-source, developers can offset this by making changes to the model.

Sharma said in the blog post that DeepSeek R1 has “undergone rigorous red teaming and safety evaluations, including automated assessments of model behavior and extensive security reviews to mitigate potential risks.”

Moreover, Microsoft’s platform has built-in content filtering and a safety evaluation system to test applications before deployment.

Chinese tech giant Alibaba released Tuesday (Jan. 28) the latest version of its AI model, Qwen2.5-Max, and said in a blog post that it outperforms DeepSeek and other top AI models on key benchmarks.

Alibaba said in the post that Qwen2.5-Max held its own against DeepSeek V3, Llama 3.1-405B, GPT-4o and Claude 3.5 Sonnet in the MMLU-Pro, GPQA-Diamond, LiveCodeBench, LiveBench and Arena-Hard benchmarks.

MMLU-Pro tests knowledge through college-level problems, GPQA-Diamond tests graduate-level capabilities, LiveCodeBench assesses coding capabilities, LiveBench tests general capabilities and Arena-Hard approximates human preferences.

Qwen2.5-Max was pre-trained on 20 trillion tokens, according to the post.

Alibaba is letting users try out its latest model through Qwen Chat, the post said. Most foundation AI models add a chatbot user interface to make it easier to use.

Qwen2.5-Max’s API is available through Alibaba Cloud, which is the largest cloud computing platform in China and Asia Pacific, and one of the largest in the world.

Meta is rolling out a new feature in Meta AI that lets it remember details when users interact with it one-on-one, according to a Monday press release. Users can tell Meta AI that they like hiking and swimming, and future responses will remember it.

Meta AI can also pick up on details within context. If a user asks for breakfast ideas and Meta AI recommends an omelet, the user can respond with a preference for vegan food only. Meta AI will remember that the user is vegan, which will inform future food recommendations, the release said.

The memory feature is coming to Facebook, Messenger and WhatsApp for iOS and Android devices in the United States and Canada. The memory feature doesn’t work in group chats. Users can delete memories at any time.

Meta AI is also getting more personalized on Facebook, Messenger and Instagram.

Users can ask Meta AI to recommend a social activity this weekend. The AI assistant will look at the user’s location in their Facebook profile, as well as what posts or videos the user looked at, to determine their interests. If the user looked at music video clips of country music, for example, Meta AI could recommend a country music show.

Google announced in a Jan. 22 blog post that Gemini Live will now be able to accept images, files and YouTube videos. The AI chatbot lets users converse with the AI assistant in a “free-flowing” manner and allows uploading of images, videos and files to further deepen the conversation.

Gemini Live is reminiscent of OpenAI’s advanced voice mode for ChatGPT, which was released in September. Both chatbots will let users talk to them in natural language on a wide range of topics.

“It is especially helpful when you want some AI-powered help to brainstorm new ideas, organize your thoughts, or simplify complex topics,” Sissie Hsiao, Google vice president and general manager of the Gemini app and speech, wrote in the blog post.

The multimodal capability already rolled out to Samsung Galaxy S24 and S25 series, as well as Google’s own Pixel 9 devices. In the coming weeks, the capability will be expanded to more Android devices. In coming months, Google said Project Astra capabilities, such as screen-sharing and live video streaming, will arrive, first in the Gemini mobile app on Android and Samsung Galaxy S25.

For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.

The post AI Tools: Microsoft Azure Adds DeepSeek; Alibaba Releases New Model; Meta Gets ‘Memories’ appeared first on PYMNTS.com.