Microsoft BitNet Releases 1-Bit Embedding Models and VibeASR Engine
Microsoft’s BitNet project has released VibeASR.cpp, a real-time multilingual ASR inference engine that uses BitNet I2_S quantization on CPUs, along with the first 1-bit embedding models at 0.6B and 270M scales. The updates promise lossless 2-bit-per-weight inference and significant embedding-prefill speedups over F16. This pushes ultra-low-bit inference closer to practical on-device speech and retrieval workloads.
GitHub