Research methods and released artifacts
Microsoft Research is Microsoft's research division, established in 1991. It works with academic collaborators and Microsoft product teams, so its outputs include papers, research software and model artifacts as well as work that feeds products. Its model releases give researchers artifacts to run as well as methods to read. [1]
The ResNet paper introduced residual learning for deep neural networks, using shortcut connections so layers learn residual functions. The authors reported experiments with substantially deeper networks, making the optimization problem itself a central part of the contribution. [2]
The division's BitNet b1.58 2B model is a downloadable language-model artifact trained on four trillion tokens and listed under the MIT license. Its ternary weights take three values; the 1.58-bit description concerns weight representation, rather than every activation and computation running at that precision. [3]