Skip to content

Microsoft Small Language Models — Phi & Aion

⚡ Microsoft Small Language Models — Phi & Aion

Section titled “⚡ Microsoft Small Language Models — Phi & Aion”

Microsoft pioneered the idea that a small model trained on the right data can outperform models many times its size. The Phi family proved this at scale. At Microsoft Build 2026, Microsoft went further — shipping Aion 1.0 directly into Windows as an on-device model that runs on CPU alone, with no cloud required.

This page covers every Microsoft SLM from the original Phi-1 in 2023 through the Aion 1.0 models in 2026.


🤖 Model📅 Released⚙️ Parameters📄 Context🎯 Focus
⚡ Phi-1Jun 20231.3B2KPython coding
⚡ Phi-1.5Sep 20231.3B2KCommon sense reasoning
⚡ Phi-2Dec 20232.7B2KLanguage + reasoning
⚡ Phi-3-miniApr 20243.8B4K / 128KGeneral language
⚡ Phi-3-smallMay 20247B8K / 128KGeneral language
⚡ Phi-3-mediumMay 202414B4K / 128KGeneral language
👁️ Phi-3-visionMay 20244.2B128KLanguage + vision
⚡ Phi-3.5-miniAug 20243.82B128KGeneral language
🧠 Phi-3.5-MoEAug 202441.9B (MoE)128KAdvanced reasoning
👁️ Phi-3.5-visionAug 20244.15B128KImage + video analysis
🧠 Phi-4Dec 202414B16KComplex reasoning + math
⚡ Phi-4-miniFeb 20253.8B128KSpeed + long context
🎭 Phi-4-multimodalFeb 20255.6B128KSpeech + vision + text
🔬 Phi-4-reasoningApr 202514B128KDeep reasoning (fine-tuned)
🪟 Aion 1.0 InstructJun 2026UndisclosedOn-device Windows tasks
🪟 Aion 1.0 PlanJun 202614B32KOn-device reasoning + agents

🔢 Phi-1 and Phi-2 — Where It All Began (2023)

Section titled “🔢 Phi-1 and Phi-2 — Where It All Began (2023)”
⚡ Phi-1 — June 2023

Phi-1 was Microsoft’s proof-of-concept that model quality is more about data quality than raw model size.

🔧 Property📋 Detail
📅 ReleasedJune 2023
⚙️ Parameters1.3 billion
📄 Context window2,048 tokens
🎯 FocusPython code generation
📜 LicenseMIT (open weights)

✅ What it showed: A 1.3B model trained on carefully curated “textbook-quality” Python data could match or beat much larger code models on Python benchmarks.

🎯 Primary use: Python code generation, code understanding, programming exercises.

⚡ Phi-1.5 — September 2023

Phi-1.5 extended the Phi-1 experiment to general language and common sense reasoning.

🔧 Property📋 Detail
📅 ReleasedSeptember 2023
⚙️ Parameters1.3 billion
📄 Context window2,048 tokens
🎯 FocusCommon sense reasoning, language
📜 LicenseMIT (open weights)

✅ What it showed: The “textbook data” approach generalised beyond code to broader reasoning tasks.

⚡ Phi-2 — December 2023

Phi-2 doubled the parameter count and demonstrated that a 2.7B model could punch well above its weight class.

🔧 Property📋 Detail
📅 ReleasedDecember 2023
⚙️ Parameters2.7 billion
📄 Context window2,048 tokens
🎯 FocusLanguage understanding + reasoning
📜 LicenseMIT (open weights)

✅ What it showed: Phi-2 outperformed Llama-2-7B and Mistral-7B on several benchmarks — a 2.7B model beating 7B models.

🎯 Primary use: Lightweight assistants, classification, summarisation, local deployment experiments.


🔢 Phi-3 Family — The Production SLM (2024)

Section titled “🔢 Phi-3 Family — The Production SLM (2024)”

The Phi-3 family was Microsoft’s first SLM family designed for real production deployment — not just research. It runs on phones, laptops, and edge devices.

⚡ Phi-3-mini — April 2024
🔧 Property📋 Detail
📅 ReleasedApril 2024
⚙️ Parameters3.8 billion
📄 Context window4K tokens (standard) · 128K tokens (long)
🎯 FocusGeneral language, reasoning
🌐 AvailabilityAzure AI · Hugging Face · ONNX · NVIDIA NIM
📜 LicenseMIT (open weights)

✅ Best for:

  • 📱 Mobile and edge device deployment
  • 🏷️ Classification and routing
  • 📝 Short-form summarisation
  • ⚡ Low-latency assistants

💡 Key result: Phi-3-mini outperforms models twice its size on several language and reasoning benchmarks.

⚡ Phi-3-small — May 2024
🔧 Property📋 Detail
📅 ReleasedMay 2024
⚙️ Parameters7 billion
📄 Context window8K tokens (standard) · 128K tokens (long)
🎯 FocusGeneral language, stronger reasoning
🌐 AvailabilityAzure AI · Hugging Face
📜 LicenseMIT (open weights)

✅ Best for:

  • 🧠 Tasks needing more reasoning than Phi-3-mini can handle
  • 📄 Longer context processing
  • 💬 Production chat assistants at moderate cost

💡 Key result: Phi-3-small outperforms GPT-3.5T on many benchmarks despite being a fraction of the size.

⚡ Phi-3-medium — May 2024
🔧 Property📋 Detail
📅 ReleasedMay 2024
⚙️ Parameters14 billion
📄 Context window4K tokens (standard) · 128K tokens (long)
🎯 FocusStrong general language and reasoning
🌐 AvailabilityAzure AI · Hugging Face
📜 LicenseMIT (open weights)

✅ Best for:

  • 🧠 Tasks requiring stronger reasoning than 7B models
  • 💻 Code understanding and explanation
  • 📄 Document analysis
  • 🏢 Enterprise RAG and assistants
👁️ Phi-3-vision — May 2024
🔧 Property📋 Detail
📅 ReleasedMay 2024
⚙️ Parameters4.2 billion
📄 Context window128K tokens
🎯 FocusLanguage + image understanding
🌐 AvailabilityAzure AI · Hugging Face
📜 LicenseMIT (open weights)

✅ Best for:

  • 👁️ Image analysis and visual Q&A
  • 🖥️ Screenshot understanding
  • 📊 Chart and diagram explanation
  • 🔤 OCR-style extraction from images

🔢 Phi-3.5 Family — Extended Capabilities (August 2024)

Section titled “🔢 Phi-3.5 Family — Extended Capabilities (August 2024)”

The Phi-3.5 series brought three new models, notably the MoE variant and improved vision capabilities.

⚡ Phi-3.5-mini — August 2024
🔧 Property📋 Detail
📅 ReleasedAugust 2024
⚙️ Parameters3.82 billion
📄 Context window128K tokens
🎯 FocusFast general language, long context
🌐 AvailabilityAzure AI · Hugging Face
📜 LicenseMIT (open weights)

✅ Best for:

  • 📄 Long document summarisation
  • 🔍 Information retrieval over long contexts
  • ⚡ Low-latency production assistants
  • 📱 Mobile and edge deployment
🧠 Phi-3.5-MoE — August 2024

The Mixture-of-Experts variant — Microsoft’s most capable Phi-3.5 model.

🔧 Property📋 Detail
📅 ReleasedAugust 2024
⚙️ Parameters41.9 billion (MoE — not all active at once)
📄 Context window128K tokens
🎯 FocusAdvanced reasoning, strong general tasks
🌐 AvailabilityAzure AI · Hugging Face
📜 LicenseMIT (open weights)

💡 What MoE means: Only a subset of the 41.9B parameters are active for any given token — giving large-model quality at lower inference cost.

✅ Best for:

  • 🧠 Complex multi-step reasoning
  • 💻 Advanced code tasks
  • 📄 Long document analysis
  • 🧩 Tasks requiring both breadth and depth
👁️ Phi-3.5-vision — August 2024
🔧 Property📋 Detail
📅 ReleasedAugust 2024
⚙️ Parameters4.15 billion
📄 Context window128K tokens
🎯 FocusImage + video analysis
🌐 AvailabilityAzure AI · Hugging Face
📜 LicenseMIT (open weights)

✅ Best for:

  • 🖼️ Image and video content understanding
  • 📊 Visual data extraction
  • 🖥️ UI and screenshot analysis
  • 🎬 Video frame-level Q&A

🔢 Phi-4 Family — Reasoning-First SLMs (December 2024 – April 2025)

Section titled “🔢 Phi-4 Family — Reasoning-First SLMs (December 2024 – April 2025)”

Phi-4 represented a major leap: reasoning ability became a first-class design goal, not just benchmark performance.

🧠 Phi-4 — December 2024
🔧 Property📋 Detail
📅 ReleasedDecember 2024
⚙️ Parameters14 billion
📄 Context window16,384 tokens
🎯 FocusComplex reasoning · mathematics · STEM
🌐 AvailabilityAzure AI · Hugging Face
📜 LicenseMIT (open weights)

✅ What makes it special:

  • 🔢 State-of-the-art on math benchmarks at 14B scale
  • 🧠 Significantly better at multi-step reasoning than Phi-3-medium
  • 💻 Strong coding and code explanation ability

💡 Example:

“Prove that the sum of two odd numbers is always even, and generalise the proof to show why the product of two odd numbers is always odd.”

⚡ Phi-4-mini — February 2025
🔧 Property📋 Detail
📅 ReleasedFebruary 26, 2025
⚙️ Parameters3.8 billion
📄 Context window128,000 tokens
🎯 FocusSpeed · long context · efficiency
🌐 AvailabilityAzure AI · Hugging Face
📜 LicenseMIT (open weights)
🔤 Vocabulary200,000 tokens (grouped-query attention)

✅ Best for:

  • 📄 Long document summarisation (up to 128K tokens)
  • ⚡ High-speed classification and extraction
  • 📱 Edge and mobile deployment
  • 🔍 Information retrieval and RAG assistants

💡 Key fact: At 3.8B parameters with a 128K context window, Phi-4-mini handles document lengths that many larger models cannot.

🎭 Phi-4-multimodal — February 2025
🔧 Property📋 Detail
📅 ReleasedFebruary 26, 2025
⚙️ Parameters5.6 billion
📄 Context window128,000 tokens
🎯 FocusSpeech + vision + text — all three at once
🌐 AvailabilityAzure AI · Hugging Face
📜 LicenseMIT (open weights)

✅ What makes it special:

  • 🎙️ Processes speech audio directly (no separate transcription step)
  • 👁️ Understands images and vision inputs
  • 💬 Handles text — all three modalities in one model

💡 Example use case:

A field technician speaks a question while pointing their phone camera at a broken circuit board → Phi-4-multimodal transcribes the speech, analyses the image, and returns a repair instruction — all on-device.

🔬 Phi-4-reasoning — April 2025
🔧 Property📋 Detail
📅 ReleasedApril 30, 2025
⚙️ Parameters14 billion
📄 Context window128,000 tokens
🎯 FocusDeep reasoning — fine-tuned from Phi-4
🌐 AvailabilityAzure AI · Hugging Face
📜 LicenseMIT (open weights)

✅ What makes it special:

  • 🔬 Fine-tuned specifically to improve chain-of-thought reasoning
  • 🔢 Excellent on maths, logic, and scientific problem solving
  • 🧠 Best Microsoft SLM for tasks where you would otherwise reach for a large reasoning model

💡 Example:

“A train leaves Station A at 60 km/h. Another leaves Station B (300 km away) at 90 km/h toward the same destination. A bird flies at 200 km/h back and forth between the trains until they meet. How far does the bird fly?”


🪟 Aion 1.0 — On-Device AI Built Into Windows (June 2026)

Section titled “🪟 Aion 1.0 — On-Device AI Built Into Windows (June 2026)”

Announced at Microsoft Build 2026, Aion 1.0 is a new generation of on-device SLMs that ship inside Windows 11. Unlike cloud models, Aion runs locally — no internet connection required, no API call, no data leaving the device.

🪟 Aion 1.0 Instruct — The Everyday Workhorse
🔧 Property📋 Detail
📅 AnnouncedJune 2, 2026 (Microsoft Build 2026)
⚙️ ParametersNot published (described as “smaller, faster, more efficient” than the previous Windows OS SLM)
🏃 HardwareRuns on CPU — no NPU or GPU required
🚀 Open sourcePlanned for Hugging Face release — July 2026
🌐 PreviewAvailable now in Microsoft Edge Insider channels

✅ What it does:

  • 📝 Summarisation of documents and web pages
  • ✏️ Text rewriting and tone adjustment
  • 🏷️ Intent detection
  • ♿ Accessibility features (in-OS)
  • 🖥️ Works on almost any Windows PC — no special hardware needed

💡 What makes it different:

Because it runs entirely on CPU, Aion 1.0 Instruct works on far more Windows devices than any previous on-device AI model — including older PCs without an NPU or dedicated GPU.

🧠 Aion 1.0 Plan — On-Device Reasoning and Agents
🔧 Property📋 Detail
📅 AnnouncedJune 2, 2026 (Microsoft Build 2026)
⚙️ Parameters14 billion
📄 Context window32,000 tokens
🎯 FocusReasoning · tool calling · agent orchestration
🏃 HardwareRuns on-device (Windows 11 built-in)
🌐 AccessShips in-box as part of Windows

✅ What it does:

  • 🧠 Reasons over user intent without sending data to the cloud
  • 🔧 Invokes local tools (file management, app control, OS functions)
  • 📁 Manages files and system tasks via agent loops
  • 🤖 Orchestrates sub-agents on the local device
  • 🔒 Fully private — no data leaves the machine

💡 Example use case:

“Organise all PDFs in my Downloads folder from the last 30 days by project name, create a subfolder for each project, and generate a markdown summary of each file.”

Aion 1.0 Plan handles this entirely on-device — no cloud, no API, no data leaving the PC.


🎯 Task✅ Best Model💡 Why
⚡ High-speed classification or routingPhi-3-mini or Phi-4-miniSmallest models with excellent classification ability
📄 Long document summarisation (up to 128K)Phi-4-mini or Phi-3.5-miniLarge context window at small model size
🔢 Maths, logic, or STEM reasoningPhi-4-reasoningFine-tuned specifically for deep reasoning
🏗️ Complex reasoning at low costPhi-3.5-MoEMoE architecture gives large-model quality
👁️ Image or screenshot analysisPhi-3.5-vision or Phi-4-multimodalBuilt-in vision understanding
🎙️ Speech + image + text in one modelPhi-4-multimodalOnly Microsoft SLM handling all three
📱 Mobile or edge deploymentPhi-3-mini, Phi-4-miniSmallest and most deployment-friendly
🏠 Fully offline, no cloud, on WindowsAion 1.0 InstructRuns on CPU, built into Windows
🤖 On-device agent and tool callingAion 1.0 Plan14B reasoning model built into Windows
💻 Code generation (general)Phi-4 or Phi-3-mediumStrong coding + reasoning combination

🏷️ Platform🤖 Models Available🔗 Notes
🔷 Azure AI FoundryFull Phi familyAPI and managed deployment
🤗 Hugging FaceFull Phi familyFree download, MIT licence
🏠 ONNX RuntimePhi-3 familyOptimised for Windows, mobile, web
🟩 NVIDIA NIMPhi-3 familyInference on NVIDIA GPUs
🔵 Intel acceleratorsPhi-3 familyOptimised inference
🪟 Windows 11 built-inAion 1.0 Instruct, Aion 1.0 PlanShips in-box, no API key required
🌐 Microsoft Edge InsiderAion 1.0 InstructPreview available now

✅ Advantage📋 Explanation
📜 Open weights (MIT)All Phi models are freely downloadable and modifiable
🏠 Local deploymentDesigned to run on phones, laptops, and edge devices
🔒 PrivacyOn-device models (Aion) keep data on the machine
💰 Cost efficiencySmaller models = lower inference cost at scale
🎓 “Textbook data” trainingQuality of training data over quantity of raw text
🌐 Broad hardware supportONNX, NVIDIA NIM, Intel — not just Azure GPUs