In the increasingly competitive landscape of artificial intelligence, Microsoft has just unveiled a technology that could radically change the way we conceive, develop and deploy AI models. It’s called BitNet b1.58 2B4T and is the first 1-bit quantised model, with a 2 billion parameter structure and a feature that could redefine the entire industry: it can run on ordinary CPUs. Yes, without the need for expensive GPUs or ultra-performing cloud infrastructures, paving the way for decentralised and ultra-accessible scenarios.
This model, unlike classical AI based on floating points or more ‘traditional’ quantisation such as 8-bit or 4-bit, uses only three possible values for its weights: -1, 0 and 1. A solution that makes it possible to drastically reduce the memory required and improve computational efficiency without compromising performance too much. In tests reported by the researchers, BitNet outperformed weight models such as those developed by Meta, Google and Alibaba on a number of benchmarks, demonstrating how extreme reduction in precision can still guarantee quality results in specific domains.
The most interesting point of this innovation, especially for those of us who follow AI and Web3 developments closely, is the possibility of imagining artificial intelligence applications embedded directly in users’ devices, without the need for centralised servers. This means that crypto wallets could host local AI assistants, Web3 gaming environments could rely on intelligent NPCs managed directly on standard devices, and on-chain analytics platforms could exploit lightweight real-time algorithms, eliminating dependencies on external data centres.
The decentralised philosophy of Web3 thus meets the possibility of a more accessible and distributed AI, less constrained by proprietary infrastructures and closer to the end user. It is a fascinating scenario, which could also redefine the concept of privacy in the relationship between artificial intelligence and personal data, with models operating directly on the local device without sending sensitive information to remote servers.
Of course, today BitNet only works through the CPU-specific bitnet.cpp framework, which is limited to certain environments. But the direction is clear, and it is only a matter of time before this type of technology becomes the new standard also for decentralised AI integrated in blockchain architectures, starting from Ethereum to more vertical Layer 2 ones such as Base.
If the future of Web3 will be increasingly AI-native, this is undoubtedly one of the most interesting news of the year. And we at BlockWorld.co.uk will be here to tell the story.
