Microsoft Research
Microsoft's in-house research organisation, founded in 1991
Profile
- Country
- United States
- Headquarters
- Redmond, Washington, USA (Microsoft campus)
- Founded
- 1991
- Founded by
- Microsoft (founder Rick Rashid; prompted by Nathan Myhrvold's 1991 plan memo)
- Type
- In-house research division of Microsoft
Microsoft Research (MSR) is a research organisation inside the American company Microsoft. It is not a separate company but a division of Microsoft. It was created after a 1991 research-lab plan memo written by Nathan Myhrvold, and Rick Rashid, a former professor at Carnegie Mellon University, set it up and led it from 1991 to 2013. Its first lab was set up on Microsoft's campus in Redmond, Washington. In March 2026, the US tech news site GeekWire reported that Igor Tsyganskiy, Microsoft's chief information security officer, had taken over from Peter Lee as executive vice president of MSR.
MSR carries out basic research in computer science and many other scientific fields. It publishes its results as papers and open code, and also uses them to improve Microsoft's products. As of October 2026, besides Redmond, it has labs in places such as Cambridge (UK), Beijing (China), Bengaluru (India), Montreal (Canada) and New York (US). Microsoft Research Asia in Beijing opened in 1998. Separately from MSR, in March 2024 Microsoft created Microsoft AI, led by Mustafa Suleyman, and gave it consumer AI products such as the AI assistant Copilot. After a reorganisation in March 2026, Copilot moved to a different organisation, and Suleyman was to focus on developing Microsoft's own AI models.
One of MSR's best-known results in AI history is ResNet, presented by MSR researchers in 2015. It is a method that lets neural networks (AI structures loosely modelled on the connections between nerve cells in the brain) keep learning well even when they become very deep, and it was used to train a network with 152 layers. With this method, the team won first place in the classification task of that year's international image-recognition competition (ILSVRC). Since 2023, MSR has been building a series of small language models called Phi. Phi-4, released in December 2024, has 14 billion parameters (the adjustable values inside a model) and is available under the MIT licence for anyone to download and use. In March 2026, MSR released a new Phi model that can also understand images, as open weights (a model published so that anyone can download and use it).
Sources for this profile (12) · as of Nov 1, 2026
- About Microsoft Research · Microsoft Research
- Labs & locations · Microsoft Research
- Rick Rashid & Founding Microsoft Research · Microsoft Research
- Microsoft Research Plan - Original memo, 1991 · Microsoft Research
- Microsoft Research · Wikipedia
- Tech Moves: Microsoft Research gets a new leader; Amazon head joins AI startup; JPMorgan exec departing · GeekWire
- Mustafa Suleyman, DeepMind and Inflection Co-founder, joins Microsoft to lead Copilot · The Official Microsoft Blog
- Deep Residual Learning for Image Recognition · Microsoft Research
- microsoft/phi-4 model card · Hugging Face (Microsoft)
- Phi-4-reasoning-vision-15B Technical Report · Microsoft Research
- Announcing Copilot leadership update · The Official Microsoft Blog
- microsoft/Phi-4-reasoning-vision-15B model card · Hugging Face (Microsoft)
AIMade by AIThis profile was written by AI (Claude) from the 12 sources above (official pages, filings, news reports).How it's made
Records 14
Newest first. Open one to see it with its source.
A case of solving, through architecture, a chronic limitation of diffusion language models: "state breaks at chunk boundaries". A substantive advance that arrived just as the observation topic "diffusion language models"
The original GraphRAG authors acknowledged the cost problem themselves and turned toward deferred graph construction
A reference implementation of the orchestrator-worker pattern
The point at which a research technique spread as a widely replicable implementation
Handles global summarization queries that local retrieval cannot answer
Justified the approach of small models deployable on phones through data curation
Solved the problem of long prompts stalling decode; 2.6–5.6× the serving capacity of vLLM
A lineage of structured indexes beyond vector similarity search
A multi-agent approach that composes agents through conversations between them and allows human-in-the-loop
A branch that applied reasoning prompting to latency optimization
Trained a 1.3B model on 7B tokens to reach 50.6% on HumanEval, demonstrating that data quality matters more than quantity
The layer that made extremely large-scale training possible without rewriting code for model parallelism
An SSD-based ANN that broke HNSW's constraint of having to hold everything in memory
Made training of very deep networks possible through residual connections. A basic component of most later architectures, including the Transformer