0.4GB 1-bit Model: Microsoft Open-Sources the First Native 1-Bit Model, Easily Runs on CPU
Microsoft's BitNet b1.58: World's 1st 1-bit LLM (0.4GB, CPU-friendly). Open-source, 2B params, rivals full-precision models!
"AI Disruption" Publication 5900 Subscriptions 20% Discount Offer Link.
Large Model Lightweighting Gets Exciting Again
Recently, Microsoft Research Asia opened-sourced the first LLM with 2 billion parameters and native 1-bit precision—BitNet b1.58 2B4T.
This model is exciting for three reasons, all embedded in its name:
1. b1.58 Quantization
The model’s parameters are limited to just three values: {-1, 0, +1}. It’s hard to imagine how knowledge is compressed into this! (Based on the information theory formula
, this model’s precision is ≈1.58 bits. Pure 1-bit quantization would have only two values.)
2. Incredibly Small Model!
With only 2B parameters, it’s a fraction of the size of full-scale models like 14B, 32B, or 617B. Thanks to extreme parameter precision compression, the model is just 0.4GB in size.
3. CPU-Targeted Inference Framework
It runs on BitNet, an open-source inference framework designed specifically for CPU architectures. Microsoft has spent 1-2 years refining this framework.