MAI-DS-R1: Microsoft’s New, More Secure and Responsive AI Model | Llm meaning software | Quick start guide to large language models pdf free download | Llm python -- tutorial | Turtles AI

Microsoft AI today announced MAI-DS-R1, a new variant of the DeepSeek R1 model with open weights, improved responsiveness and security. Deployed via Azure AI Foundry and Hugging Face, it delivers high performance on sensitive topics.

Key Points:

• Improved responsiveness: MAI-DS-R1 correctly answers 99.3% of prompts on blocked topics, outperforming DeepSeek R1.
• Enhanced security: Significant reduction in malicious content, with microattack success rate cut in half.
• Stable performance: Maintains high reasoning capabilities on general knowledge, math, and coding benchmarks.
• Accessibility: Available as an open-weight model and via API on Azure AI Foundry and Hugging Face.

Microsoft AI today announced MAI-DS-R1, a new variant of the DeepSeek R1 model designed to improve handling of sensitive topics and response security. Deployed via Azure AI Foundry and Hugging Face, MAI-DS-R1 was post-trained on approximately 350,000 examples of blocked topics, using strategies such as keyword collection and filtering, translation of questions into multiple languages, and the use of internal models to bootstrap answers. Additionally, 110,000 examples of security and non-compliance from the Tulu3 SFT dataset were included.

Evaluations showed that MAI-DS-R1 correctly answered 99.3% of the blocked topic prompts, outperforming DeepSeek R1 by 2.2x and matching Perplexity R1-1776. In terms of satisfaction, it scored higher than both models. On the security front, MAI-DS-R1 more than halves the average success rate of microattacks, consistently outperforming DeepSeek R1 and R1-1776 in nearly all functional and semantic categories, as measured by HarmBench evaluations.

Despite these improvements, MAI-DS-R1 maintains the same reasoning capabilities across the general knowledge, reasoning, math, and coding benchmarks. The model is available as an open-weight build and via an Azure-hosted API. To access the model sheet and weights, visit Hugging Face. Additionally, a highly optimized inference runtime is now available via the Azure Foundry-hosted API.

This development represents a significant step in the evolution of AI models, providing more secure and responsive tools for researchers, developers, and enterprises.