Topics
Latest
AI
Amazon
Image Credits:Aleksander Kalka/NurPhoto / Getty Images
Apps
Biotech & Health
clime
Cloud Computing
Commerce
Crypto
Enterprise
EVs
Fintech
Fundraising
widget
Gaming
Government & Policy
computer hardware
Layoffs
Media & Entertainment
Meta
Microsoft
seclusion
Robotics
security measures
Social
blank
startup
TikTok
transit
speculation
More from TechCrunch
Events
Startup Battlefield
StrictlyVC
Podcasts
Videos
Partner Content
TechCrunch Brand Studio
Crunchboard
Contact Us
Microsoft researchers take they ’ve developed the largest - ordered series 1 - bit AI model , also love as a “ bitnet , ” to particular date . promise BitNet b1.58 2B4 T , it’sopenly availableunder an MIT licence and can run on CPUs , include Apple ’s M2 .
Bitnets are essentially constrict models plan to campaign on lightweight hardware . In stock fashion model , weights , the values that define the inner social system of a model , are oftenquantized so the models perform well on a wide range of machines . Quantizing the weight unit lowers the number of bit — the smallest units a information processing system can process — needed to present those free weight , enable example to unravel on french-fried potatoes with less remembering , faster .
Bitnets quantise weight into just three time value : -1 , 0 , and 1 . In hypothesis , that makes them far more memory- and computing - efficient than most models today .
The Microsoft research worker say that BitNet b1.58 2B4 T is the first bitnet with 2 billion parametric quantity , “ parameters ” being largely synonymous with “ weights . ” Trained on a dataset of 4 trillion tokens — equivalent to about 33 million books , by one estimation — BitNet b1.58 2B4 T exceed traditional example of similar size , the researchers claim .
BitNet b1.58 2B4 T does n’t sweep the floor with rival 2 billion - argument modeling , to be clear , but it apparently holds its own . agree to the researchers ’ examination , the theoretical account surpasses Meta ’s Llama 3.2 1B , Google ’s Gemma 3 1B , and Alibaba ’s Qwen 2.5 1.5B on benchmarks including GSM8 K ( a compendium of level - school - horizontal surface math problems ) and PIQA ( which prove forcible commonsense reasoning attainment ) .
Perhaps more imposingly , BitNet b1.58 2B4 T is speedier than other models of its size — in some cases , twice the velocity — while using a fraction of the memory .
There is a catch , however .
Join us at TechCrunch Sessions: AI
Exhibit at TechCrunch Sessions: AI
Achieving that performance requires using Microsoft ’s custom theoretical account , bitnet.cpp , which only works with certain computer hardware at the moment . abstracted from the list of supported chip are GPUs , which dominate the AI substructure landscape .
That ’s all to say that bitnets may hold promise , in particular for imagination - constrained equipment . But compatibility is — and will likely stay on — a big sticking point .
