Microsoft

public

usa · Founded 1975 · IPO 1986 (NASDAQ: MSFT) · Market cap: $3T as of 2026-04

Microsoft Research (MSR), founded in 1991, is one of the world's largest industrial research labs with 1,000+ researchers across Redmond, Cambridge, Beijing, Bangalore, NYC, and Montreal. Microsoft is also OpenAI's largest investor (~27% equity) and integrates OpenAI models into Copilot products.

Microsoft's own model research focuses on small language models: the Phi series demonstrated that carefully curated synthetic data can make small models (1.3B–14B) competitive with models 10–100x larger. Phi-1 (2023, "Textbooks Are All You Need") launched the series, and Phi-4 (2024, 14B) surpasses its teacher model on STEM benchmarks.

MSR produced several foundational techniques: LoRA (Low-Rank Adaptation, 2021) became the dominant parameter-efficient fine-tuning method, BitNet (2023) pioneered 1-bit Transformers, and DeepSpeed (42K stars) is critical distributed training infrastructure used to train BLOOM-176B and other frontier models.

Website Wikipedia HuggingFace GitHub OpenRouter
frontieropen-weightinfrastructurereasoning

People

News

Outputs (14)

Horizon Length in LLM Agent Training

paper

"On Training Large Language Models for Long-Horizon Tasks: An Empirical Study of Horizon Length." Studies how task horizon length affects RL training…

Harrier

model

Open-source multilingual text embedding model family optimized for agentic AI systems. Decoder-only architecture with last-token pooling and L2 normal…

MAI Foundation Models

model

Microsoft's first major foundation model push, releasing three MAI models in Azure Foundry — a direct shot at OpenAI and Google dependencies. MA…

Speech LLMs as Contextual Reasoning Transcribers

paper

Introduces Chain-of-Thought ASR (CoT-ASR), enabling LLMs to analyze speech input and generate contextual reasoning before performing transcription. Us…

Nature Language Model (NatureLM)

model

Sequence-based science foundation model from Microsoft Research AI for Science that unifies small molecules, materials, proteins, DNA, and RNA for tex…

Phi-4

model

14B parameter dense Transformer with emphasis on synthetic data for complex reasoning. Surpasses its teacher model on STEM benchmarks. Also released a…

RedStone

dataset

Scalable pipeline for curating LLM pretraining data from Common Crawl. Produces ~3.17 trillion tokens across four data types: general web text, code,…

BioT5+

paper

"Towards Generalized Biological Understanding with IUPAC Integration and Multi-task Tuning." Builds on BioT5 by incorporating IUPAC nomenclature, exte…

BitNet

paper

"BitNet: Scaling 1-Bit Transformers for Large Language Models." Replaces linear layers with BitLinear using binary {-1, +1} weights. The follow-up Bit…

ToRA: Tool-Integrated Reasoning Agent

paper

Introduces tool-integrated reasoning agents that interleave natural language reasoning with program-based tool use (computation libraries, symbolic so…

Phi-1 ("Textbooks Are All You Need")

model

"Textbooks Are All You Need." 1.3B parameter code-focused model trained on high-quality "textbook-like" synthetic data. Achieved 50.6% on HumanEval de…

DeepSpeed-MoE

paper

End-to-end MoE training and inference system that addressed the two main barriers to Mixture-of-Experts adoption: massive model size and high inferenc…

LoRA (Low-Rank Adaptation)

paper

"LoRA: Low-Rank Adaptation of Large Language Models." Freezes pre-trained weights and injects trainable low-rank decomposition matrices into each Tran…

DeepSpeed

library

Distributed training and inference library. Introduced ZeRO (Zero Redundancy Optimizer) for memory-efficient data parallelism, 3D parallelism (data +…