Foundry Friday AMA · Jun 20, 2025 · Advanced Reasoning

[nitya]

⁉️ Have questions on reasoning models? Add them here! We'll post a recap after the AMA on this thread

---

AMA On Advanced Reasoning:
This is part of the #ModelMondays project where we invite a subject matter expert to our Monday livestream to demo a key model or model-related technology - then have them do an AMA on Friday on the Azure AI Foundry Discord to take questions and have a more open-ended conversation. See: #FoundryFridayAMA Series for the full schedule of AMA events.

This week we'll focus on Advanced Reasoning Models with our Subject Matter Expert: Marlene Mhangami

1️⃣ | Register for the Friday AMA - 1:30pm ET
2️⃣ | Watch the Monday Livestream- 1:30pm ET
3️⃣ | Learn more About Model Mondays - Season 1 Recaps + Season 2 Schedule

---

[nitya]

Monday Livestream

Missed the Model Mondays livestream? Catch up on the replay!. Here is what we covered:

1. Highlights - Top 5 News Items in 5 mins.
2. Spotlight - Advanced Reasoning Models with Marlene M.
3. Community Discussion - Q&A from chat

Slide Deck

The slide deck contains key takeaways from the session with handy QR codes to quickly look up the relevant news items, models and learning resources we mentioned. Download it here

Here are a subset of the models that we mentioned on the show today

Chat Q&A

1. ​Q: What's the Python library for visuals? Answer: Textual
2. Q: ​​I live in c# Whats all this python? is it worth learning? Answer (Yes, absolutely! It's a great ecosystem of libraries and worth learning) (It's easy to learn with AI help)
3. Q: What is the link to the code for the example shown? Answer:https://aka.ms/build/lab331

[nitya]

Learning Resources

Lab 333 - Introduction to Reasoning Models (Beginner-friendly)

---

Lab 331 - Build a Deep Researcher with DeepSeek-R1 and LangGraph

---

Deep Researcher Code Sample - Application used in lab above

[nitya]

This post was generated with AI help and human revision & review. To learn more about our motivation and workflows, please refer to this document in our website.

Attendees

• Nitya Narasimhan, PhD
• Lee Stott [Microsoft]
• Marlene Mhangami [Microsoft]
• Sharmila Chockalingam [Microsoft]
• Community members from various locations

Locations

• New York, USA
• Dallas, USA
• Chennai, India
• Portugal
• Vancouver, Canada
• London, UK
• Egypt
• Oregon, USA

Core Topics Discussed

• Large Reasoning Models (LRMs) and their applications
• GitHub Copilot Agent Mode and supported models
• Deep Researcher using DeepSeek-R1 and LangChain
• Multi-agent architectures and reasoning in AI
• Knowledge graphs and RAG to reduce hallucinations
• Use cases for reasoning models in coding, business, and research

Key Takeaways

1. LRMs are advancing AI's ability to reason and solve complex problems.
2. GitHub Copilot Agent Mode supports multiple top-tier models and custom integrations.
3. Deep Researcher tools like DeepSeek-R1 and LangChain enable advanced research workflows.
4. Multi-agent systems and knowledge graphs can help reduce hallucinations and improve reliability.
5. Community members are actively experimenting with reasoning models for diverse use cases.

Key Resources

• Model Mondays Playlist
• Illusion of Thinking (Apple ML Research)
• Anthropic Multi-Agent Research System
• Programming with GitHub Copilot Agent Mode
• OpenAI Deep Research
• Deep Researcher Lab (Build)
• Function Calling with Voice using GPT-4o
• Lab333: Evaluate Reasoning Models
• LazyGraphRAG (Microsoft Research)

Microsoft Resources

• Azure OpenAI reasoning models
• How to use reasoning models with Azure AI Foundry
• Transparency Note for Azure OpenAI Service
• Overview: Use o1 reasoning model in prompt builder

[nitya]

Model Mondays S2:E01 – Advanced Reasoning

This post was generated with AI help and human revision & review. To learn more about our motivation and workflows, please refer to this document in our website

Click Banner To Watch Replay

Abstract

Season 2 of Model Mondays kicks off with a deep dive into advanced reasoning models. The session covers the latest updates from Azure AI Foundry, introduces new models and features, and spotlights how reasoning models can be leveraged for high-quality, step-by-step problem solving. Marlene Mahungami demonstrates practical applications of reasoning models using DeepSeek and LangChain, highlighting their strengths, limitations, and best practices for integrating them into real-world AI solutions.

News Highlights

• Azure AI Foundry now offers a broader range of models, including DeepSeek, XAI, Meta, and Mistral, under Microsoft’s service terms.
• The new Video Playground in Foundry enables prompt-to-video generation with exportable code and enterprise-ready pipelines.
• Model Router allows automatic selection of the best model for a given use case, optimizing for cost and performance.
• Grok 3 and Grok 3 Mini models are now available, offering long context and advanced reasoning capabilities.
• DeepSeek R1-10528 introduces improved reasoning, hallucination blocking, and is ideal for agents and co-pilots.

Tech Spotlight

Marlene Mahungami presents a hands-on lab demonstrating how to build a deep researcher using reasoning models with LangChain and DeepSeek. She explains that reasoning models, trained with reinforcement learning, provide more detailed and higher-quality outputs than standard chat models. The demo showcases how to separate the model’s thought process (enclosed in tags) from the final user-facing response, and how to customize output formats (e.g., tables, bullet points) via system prompts. Marlene also discusses integrating external tools and context, such as web search, to enhance the model’s knowledge and relevance, and highlights the importance of evaluation and self-reflection in model outputs.

The conversation addresses the trade-offs between different reasoning models (OpenAI, DeepSeek, Grok), considerations for local vs. cloud deployment, and the value of open-source options for privacy and cost savings. The session concludes with a Q&A on evaluation strategies, prompt consistency, and the mechanics of reasoning content in model outputs.

Takeaways

• Reasoning models deliver higher-quality, step-by-step outputs but may have longer response times.
• Separating the model’s reasoning from the final answer improves user experience and application design.
• Customizing prompts and output formats enables tailored, production-ready solutions.
• Integrating external data sources enhances the relevance and accuracy of reasoning models.
• Continuous evaluation and self-reflection are essential for maintaining quality and consistency in AI applications.

Related Resources

• How to use reasoning models with Azure AI Foundry Models (Python)
• How to use reasoning models with Azure AI Foundry Models (Java)
• How to use reasoning models with Azure AI Foundry Models (C#)
• How to use reasoning models with Azure AI Foundry Models (JavaScript)
• Azure AI Foundry Model Inference Documentation