BitNet, Microsoft เปิดซอร์สเฟรมเวิร์กที่รัน LLM 100B พารามิเตอร์บน CPU ตัวเดียว

Microsoft has released the open-source BitNet framework, enabling the running of a 100-billion-parameter language model (LLM) on a single CPU without the need for GPUs, accelerators, or expensive hardware. The BitNet framework specializes in 1-bit LLMs, which utilize only three weight values: -1, 0, and +1, a technique called ternary quantization or 1.58-bit quantization.

This contrasts with traditional LLMs that employ 32-bit or 16-bit floating-point numbers, which require GPUs for large models. While the technical report suggests a 100B BitNet model, the actual release is the smaller BitNet-b1.58-2B-4T, containing 2.4 billion parameters trained on 4 trillion tokens. The framework demonstrates significant performance improvements over competing models like llama.cpp, with speedups ranging from 2.37x to 6.17x on x86 CPUs and 1.37x to 5.07x on ARM CPUs.

Energy consumption is reduced between 71.9% and 82.2% on x86 CPUs. BitNet is open-source under the MIT License and available on GitHub, with 39,900+ stars and 3,700+ forks. The framework is suitable for edge computing, privacy-first applications, IoT devices, and offline scenarios where large-scale, high-speed models are unnecessary.