Microsoft released bitnet-embedding-0.6b and a smaller bitnet-embedding-270m, text embedding models built using BitNet's 1-bit (technically ternary, meaning each weight is restricted to -1, 0, or 1) quantization scheme rather than the 16- or 32-bit floating point weights most embedding models use. Embedding models convert text into numeric vectors for search, retrieval-augmented generation, clustering, and deduplication, and they are usually the highest-QPS component in a RAG pipeline because every query and every document chunk has to pass through one. Compressing the weights to roughly 2 bits per parameter cuts memory footprint dramatically and, more importantly for real deployments, delivers a measured 1.42x to 2.28x speedup over standard 16-bit inference on ordinary x86 CPUs, without requiring a GPU at all. That matters because embedding is frequently the part of a RAG or search pipeline that teams end up over-provisioning GPUs for simply to keep up with query volume, and a model that can hit competitive embedding quality on CPU inference changes that cost equation substantially for high-volume, latency-sensitive retrieval workloads. The models are published on Hugging Face with ready integrations for llama.cpp, Ollama, LM Studio, and Docker Model Runner, so they slot into existing local-inference tooling rather than requiring a bespoke serving stack. Microsoft is explicit that the models are aimed at research and further evaluation rather than being drop-in production-ready today, but the release is a concrete data point that ternary-weight quantization, which BitNet has been developing for full language models, works for the embedding use case specifically, not just for generative decoding.