Topics
Latest
AI
Amazon
Image Credits:Larysa Amosova via Getty
Apps
Biotech & Health
Climate
Image Credits:Larysa Amosova via Getty
Cloud Computing
Commerce
Crypto
Enterprise
EVs
Fintech
Fundraising
Gadgets
back
Government & Policy
computer hardware
Layoffs
Media & Entertainment
Meta
Microsoft
Privacy
Robotics
protection
societal
Space
Startups
TikTok
exile
Venture
More from TechCrunch
Events
Startup Battlefield
StrictlyVC
Podcasts
telecasting
Partner Content
TechCrunch Brand Studio
Crunchboard
Contact Us
Meta hasannouncedthe novel addition to itsLlamafamily of reproductive AI models : Llama 3.3 70B.
In aposton X , Ahmad Al - Dahle , VP of procreative AI at Meta , said that the text - only Llama 3.3 70B redeem the performance of Meta ’s with child Llama modelling , Llama 3.1 405B , at lower cost .
“ By leverage the latest onward motion in post - training proficiency … this model better nucleus operation at a significantly scurvy cost , ” Al - Dahle pen .
Al - Dahle publish a chart showing Llama 3.3 70B surmount Google ’s Gemini 1.5 Pro , OpenAI ’s GPT-4o , and Amazon ’s new free Nova Pro on a number of industry benchmark , include MMLU , which evaluates a model ’s power to understand words . Via email , a Meta spokesperson said that the model should cede improvement in areas like maths , general knowledge , instruction abide by , and app use .
introduce Llama 3.3 — a new 70B model that have the performance of our 405B mannequin but is easier & more cost - efficient to fly the coop . By leveraging the latest progress in post - training techniques including online preference optimisation , this poser better core performance at…pic.twitter.com/6oQ7b3Yuzc
— Ahmad Al - Dahle ( @Ahmad_Al_Dahle)December 6 , 2024
Llama 3.3 70B , which is available fordownloadfrom the AI dev political program Hugging Face and other sources , including the official Llamawebsite , is Meta ’s latest play to eclipse the AI playing field with “ open ” manakin that can be used and market for a kitchen stove of applications .
Meta ’s termsconstrainhow sure developer can use Llama models ; platforms with more than 700 million monthly users must request a especial license . But for many , it ’s incorporeal that Llama model are n’t “ open ” in the strictest sense . Case in item , Llama has racked up more than 650 million downloads , concord to Meta .
Join us at TechCrunch Sessions: AI
Exhibit at TechCrunch Sessions: AI
Meta has leveraged Llama internally as well . Meta AI , the companionship ’s AI assistant , which is power only by Llama models , now has nearly 600 million monthly participating users , perMeta chief operating officer Mark Zuckerberg . Zuckerberg claims that Meta AI is on track to be the most - used AI help in the humanity .
For Meta , the open nature of Llama has been a blessing and a oath . In November , areportalleged that Formosan military researchers had used a Llama framework to educate a defence chatbot . Meta responded by do its Llama modelsavailable to U.S. defense contractors .
Meta has also voice business about its ability to comply withthe AI Act , the EU law that establish a regulatory framework for AI , calling the law ’s implementation “ too unpredictable ” for its undefendable release strategy . A related emergence for the fellowship are provisions in the GDPR , the EU ’s concealment police force , pertaining to AI education . Meta trains AI exemplar on the public information of Instagram and Facebook users who have n’t opted out — datum that in Europe is open to GDPR guarantees .
EU regulators sooner this year requested that Meta halt breeding on European exploiter data while they assessed the ship’s company ’s GDPR compliance . Meta relented , while at the same time indorse anopen lettercalling for “ a modern interpretation ” of GDPR that does n’t “ rule out procession . ”
Meta , not immune to the technical challenges other AI labs areencountering , is ramping up its figure substructure to cultivate and serve succeeding generations of Llama . The caller announced Wednesday that it would build a $ 10 billion AI datum heart in Louisiana — the largest AI data point center Meta has ever built .
Zuckerbergsaidon Meta ’s Q4 earnings call in August that to direct the next major set of Llama models , Llama 4 , the company will require 10x more compute than what was postulate to train Llama 3 . Meta has pimp a cluster of more than 100,000 Nvidia GPUs for good example development , rivaling the resourcefulness of competition like xAI .
grooming productive AI modeling is a costly business . Meta ’s capital letter expenditures arise about 33 % to $ 8.5 billion in Q2 2024 , up from $ 6.4 billion a class in the first place , drive by investments in server , data point mall , and meshwork infrastructure .
