Foundry Friday AMA · Jul 03, 2025 · SLMs & Reasoning (Phi-4 Reasoning)

[nitya]

AMA on SLMs and Reasoning

This is part of the #ModelMondays series where we put the spotlight on a new model-related topic each week.

🌟🌟 See #54 for the full Foundry Fridays AMA schedule 🌟🌟

---

Event Details

How can you bring advanced reasoning to resource-constrained devices? This session explores the latest in Small Language Models (SLMs) like Phi-4, which are redefining what’s possible for agentic apps. Mojan Javaheripi will discuss how SLMs leverage inference-time scaling and chain-of-thought reasoning to deliver powerful results on smaller hardware. Discover use cases, deployment strategies, and how SLMs are making AI more accessible and efficient for everyone.

• 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

Related Resources

• Concepts - Small and large language models
• Phi-4 and Phi-4 Reasoning Models in Azure AI Foundry
• How to use reasoning models with Azure AI Foundry

[nitya]

Model Mondays S2:E03 – SLMs and Reasoning (Livestream Recap)

Watch The Replay

Abstract

This episode of Model Mondays spotlights Small Language Models (SLMs) and the Phi-4 reasoning model family from Microsoft Research. Mojan Javaheripi provides an in-depth look at the design, training, and real-world applications of the Phi-4 reasoning models, highlighting their efficiency, performance, and adaptability for a wide range of AI tasks.

News Highlights

• Azure AI Foundry introduces safety leaderboards for evaluating models on trustworthiness, jailbreak resistance, and prompt injection resilience.
• Foundry Local is now in public preview, enabling containerized, air-gapped deployment of models for edge and compliance use cases.
• MCP support in Azure AI Agent Service streamlines context routing and model selection for multi-model applications.
• Fine-tuning capabilities have been expanded, including reinforcement and supervised fine-tuning for various models.
• Microsoft is recognized as a leader in the 2025 Gartner Magic Quadrant for data science and machine learning platforms.

Tech Spotlight

Mojan Javaheripi explains the motivation behind SLMs: bridging the gap between large frontier models and smaller, more efficient models that can run on commodity hardware. The Phi-4 reasoning models are designed for explicit, step-by-step problem solving, with a focus on explainability and logical decomposition.

Mojan details the training pipeline, including data curation, supervised fine-tuning, and reinforcement learning, and shares benchmarking results that show Phi-4 models outperforming much larger models in math, science, and coding tasks. The discussion covers community adoption, quantization for edge devices, and the importance of fine-tuning for domain-specific applications. Real-world use cases include intelligent tutoring, autonomous agents, code generation, and strategic planning.

Takeaways

• Phi-4 reasoning models deliver state-of-the-art reasoning and efficiency at a fraction of the size of frontier models.
• SLMs are ideal for edge, local, and resource-constrained deployments, with strong support for quantization and customization.
• Fine-tuning and reinforcement learning enable rapid adaptation to new domains and tasks.
• Open-source availability and community contributions accelerate innovation and adoption.
• Responsible AI, safety, and explainability are core to the Phi-4 model family’s design and deployment.

Related Resources

• Azure AI Foundry Models available for serverless API deployment
• How to use reasoning models with Azure AI Foundry Models (Python)
• Featured models of Azure AI model catalog

Azure AI Foundry Model Cards for Phi-4-Reasoning:

• Phi-4-Reasoning
• Phi-4-mini-reasoning
• Phi-4-reasoning-plus
• Phi-4-Reasoning-Onnx
• Phi-4-mini-reasoning-Onnx

Hugging Face Model Cards for Phi-4-Reasoning:

• microsoft/Phi-4-reasoning-plus
• microsoft/Phi-4-mini-reasoning
• unsloth/Phi-4-reasoning-plus-GGUF
• microsoft/Phi-4-reasoning
• unsloth/Phi-4-mini-reasoning-GGUF

Explore Quantized Versions:

1. Using Unsloth - with GGUF, 4-bit and 16-bit formats
2. Use AI Toolkit in VS Code - 4-bit quantized, NPU optimized version for Windows Snapdragon devices

[nitya]

Episode 3 Recap: SLMs and 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

Attendees

• Lee Stott [Microsoft]
• Mojan Javaheripi [Microsoft]
• Nitya Narasimhan, PhD [Microsoft]
• Community members from various regions

Locations

• India (Kolkata, general)
• South Africa
• Sri Lanka
• Canada
• UK
• New York, USA

Core Topics Discussed

• Phi-4 reasoning models and SLMs for agentic applications
• Using Phi-4 with Ollama and local hosting
• Model catalog and resources for SLMs
• Fine-tuning, quantized versions, and LORA for prompts
• Benchmarks for agentic tool use (BFCL)
• Foundry-Local and observability in Azure AI Foundry
• Use cases: code generation, optimization problems, transparency

Key Takeaways

1. Phi-4 reasoning models are redefining SLMs for agentic and resource-constrained scenarios.
2. Community members are experimenting with local hosting and fine-tuning of SLMs.
3. Microsoft provides extensive resources, tutorials, and benchmarks for SLMs and agentic evaluation.
4. Observability and transparency are key for deploying and evaluating SLMs in production.
5. The Phi Cookbook and Foundry-Local are valuable resources for developers.

Key Resources

• Model Mondays Playlist
• Phi-4 Reasoning (Hugging Face)
• Phi-4 Reasoning Plus (Hugging Face)
• Phi-4 Reasoning Model Catalog
• Phi-4 Reasoning on GitHub Models
• Phi Cookbook
• BFCL Benchmark
• Foundry-Local
• Foundry-Local Docs
• Observability Docs

Microsoft Resources

• Phi-4 Reasoning Model Overview
• How to use reasoning models with Azure AI Foundry
• Featured models of Azure AI model catalog
• Explore Azure AI Foundry Models