Our long-term mission is to advance AI for humanity. Currently, we are dedicated to advancing the frontier of AI research and building next-generation AI models.

Building Next-Generation AI Models

• 12/25: VibeVoice-Realtime-0.5B
• 08/25: VibeVoice
• 04/25: BitNet b1.58 2B4T
• more coming

Frontier AI Research

We published and open-sourced our research and models on Foundation Models and General AI.

• Model Architecture: BitNet (v1 | b1.58 | a4.8 | v2) / 1-bit LLMs, YOCO / Decoder-Decoder Architecture, DIFF / Differential Transformer, ReSa / Rectified Sparse Attention, MH-MoE / 1-bit MoE, RetNet, LongNet, DeepNet
• Learning Paradigm: TPT / Thinking Augmented Pre-Training, RPT / Reinforcement Pre-Training, RRM / Reward Reasoning Model, Adaptive Thinking Model
• RL in LLMs: GMPO, OPO
• Multimodality: VibeVoice, LatentLM / Latent Language Modeling, (M)VoT, Kosmos-*, VALL-E*
• Multiagent: Agentic Organization / Asynchronous Thinking
• Distillation: GAD (Black-Box On-Policy Distillation), MiniLLM (On-Policy Distillation)
• Synthetic Data: QUEST, SynthLLM, GLAN
• System & Hardware: BitNet.cpp / 1-bit AI Infra


• OSS Models on GitHub/HuggingFace: VibeVoice-Realtime-0.5B, VibeVoice, BitNet b1.58 2B4T, (m)E5, Kosmos-2, WavLM, LaytoutLM(-1/2/3), InfoXLM, MiniLM

• Large-Scale Pre-trained Models: BEiT-3, BEiT, WavLM, LayoutLM*, XLM-E, InfoXLM, MiniLM*, UniLM

• OSS Projects
• microsoft/VibeVoice: Open-Source Frontier Voice AI
• microsoft/BitNet: 1-bit AI Infra: Official inference framework for 1-bit LLMs
• microsoft/unilm: Large-scale Self-supervised Pre-training Across Tasks, Languages, and Modalities

We are hiring at all levels (including FTE researchers and interns)! If you are interested in working with us on Foundation Models and General AI, NLP, Speech, and Multimodal AI, please send your resume to fuwei@microsoft.com.