Microsoft Research (Kaiming He et al.) published Deep Residual Learning for Image Recognition (ResNet), introducing residual skip connections ($y = F(x) + x$) that allowed gradient flow across 152-layer CNNs and later became an essential structural primitive inside every Transformer layer.
Microsoft AI
Also known as: Microsoft Research · MSR · MAI · Azure AI · Microsoft Foundry
Plain English
Microsoft's AI research and product organization—spanning foundational Microsoft Research breakthroughs (ResNet, DeepSpeed, LoRA, and the open-weight Phi small language models), Azure AI supercomputing and custom Maia silicon, and the first-party MAI frontier model family powering Copilot and GitHub Copilot.
Technical Overview
Microsoft operates a dual open-and-frontier AI stack: Microsoft Research pioneered residual networks, ZeRO Distributed Training (DeepSpeed), Low-Rank Adaptation (LoRA), and synthetic-data-trained Phi SLMs (1.3B–15B), while the Microsoft AI (MAI) division trains proprietary sparse MoE reasoning, coding, speech, image, and cybersecurity models (MAI-Thinking-1, MAI Code 2, MAI Realtime 1) served on Azure and custom Maia accelerators.
Timeline (12)
2015
Paper2019
EcosystemMicrosoft invested $1 billion in OpenAI (later expanded by ~$10 billion in January 2023) to co-design large-scale Azure AI supercomputing clusters and integrate frontier GPT models across GitHub Copilot, Bing, Microsoft 365, and Azure OpenAI Service.
2020
ReleaseMicrosoft open-sourced DeepSpeed built around the ZeRO (Zero Redundancy Optimizer) paper, partitioning optimizer states, gradients, and parameters across data-parallel GPUs to eliminate memory redundancy in multi-hundred-billion-parameter Distributed Training.
2021
PaperMicrosoft researchers (Edward Hu et al.) introduced LoRA (Low-Rank Adaptation), freezing pretrained model weights and injecting trainable low-rank matrices into Transformer layers—cutting fine-tuning VRAM up to 3x with zero inference latency penalty.
2023
PaperMicrosoft Research published Textbooks Are All You Need, introducing Phi-1 (1.3B) and demonstrating that curating "textbook-quality" Synthetic Data & Self-Play allows small language models to match or beat models 10x larger on Python coding benchmarks.
- Release
At Microsoft Ignite 2023, Microsoft unveiled its first custom in-house AI accelerator silicon, Azure Maia 100 (a 5nm tensor chip paired with rack-level liquid cooling and 1.8 TB/s HBM memory for cloud training and inference), alongside the 128-core ARM-based Azure Cobalt 100 CPU.
2024
MilestoneMicrosoft established the Microsoft AI (MAI) organization and appointed DeepMind co-founder and Inflection AI CEO Mustafa Suleyman as EVP and CEO (with Karén Simonyan as Chief Scientist), unifying Copilot product engineering with a mandate to train proprietary first-party frontier models.
- Release
Following the April 2024 rollout of the mobile-ready Phi-3 series (3.8B–14B), Microsoft Research released Phi-4 (14B) under an MIT license, combining multi-agent Synthetic Data & Self-Play curricula with pivotal-token DPO to surpass much larger models on mathematical and STEM reasoning.
2025
ReleaseMicrosoft AI unveiled its first two in-house foundation models trained independently of OpenAI: MAI-Voice-1 for low-latency expressive speech generation and MAI-1-preview, a sparse Mixture-of-Experts model trained across ~15,000 NVIDIA H100 GPUs.
- Milestone
Microsoft AI launched the MAI Superintelligence Team under Mustafa Suleyman and Karén Simonyan and published its Towards Humanist Superintelligence charter prioritizing controllable, domain-specific AI across software engineering, medical diagnostics, and clean energy—followed on November 18 by a $30 billion Azure compute partnership with Anthropic and NVIDIA.
2026
ReleaseAt Microsoft Build 2026, Microsoft AI launched seven first-party MAI models led by MAI-Thinking-1 (a ~1T-total / 35B-active sparse MoE reasoning model trained from scratch without third-party distillation), alongside MAI-Code-1 Flash, MAI-Cyber-1 Flash, MAI-Image-2, MAI-Voice-2, and MAI-Transcribe-1 in Microsoft Foundry.
- Release
At Microsoft Ignite 2026, Microsoft AI introduced MAI Code 2 and MAI Code 2 Flash for GitHub Copilot and Microsoft Foundry, MAI Cyber 2 Flash (1M-token cybersecurity defense), MAI Realtime 1 (native speech-to-speech at $0.01/minute), MAI Image 2.7, MAI Transcribe 2, and MAI Licensed Weights enabling enterprises to customize and self-host frontier MAI checkpoints inside their own Azure environments.
See also (16)
Open full graphMicrosoft AI
11 more connections
- LoRA & QLoRAOutgoing · Enables
- OpenAIOutgoing · Alternative to
- US AI LabsIncoming · Enables
- Convolutional Neural NetworksOutgoing · Mentions
- DPOOutgoing · Mentions
- GPTOutgoing · Mentions
- Mixture of Experts (MoE)Outgoing · Mentions
- NVIDIAOutgoing · Mentions
- Reasoning & Test-Time ComputeOutgoing · Mentions
- Synthetic Data & Self-PlayOutgoing · Mentions
- TransformerOutgoing · Mentions