Topics
up-to-the-minute
AI
Amazon
Image Credits:piranka / Getty Images
Apps
Biotech & Health
Climate
Image Credits:piranka / Getty Images
Cloud Computing
DoC
Crypto
endeavor
EVs
Fintech
fundraise
Gadgets
Gaming
Government & Policy
Hardware
layoff
Media & Entertainment
Meta
Microsoft
concealment
Robotics
Security
Social
outer space
Startups
TikTok
shipping
Venture
More from TechCrunch
Events
Startup Battlefield
StrictlyVC
Podcasts
Videos
Partner Content
TechCrunch Brand Studio
Crunchboard
adjoin Us
‘ Tis the week for small AI framework , it seems .
Nonprofit AI research institute Ai2 on ThursdayreleasedOlmo 2 1B , a 1 - billion - parameter model that Ai2 claims beats similarly - sized model from Google , Meta and Alibaba on several bench mark . parameter , sometimes refer to as weighting , are the home component part of a model that guide its behavior .
Olmo 2 1B is available under a permissive Apache 2.0 license on AI dev platform Hugging Face . Unlike most model , Olmo 2 1B can be double from slit , as Ai2 has provided the code and information set ( Olmo - mix-1124andDolmino - mix-1124 ) used to develop it .
Small exemplar might not be as subject as their behemoth counterparts , but significantly , they do n’t require beefy hardware to run . That piddle them much more approachable for developers and hobbyists contending with the limitations of lower - end hardware and consumer motorcar .
There ’s been a raft of small model launches over the past few days , from Microsoft’sPhi 4 logical thinking familytoQwen ’s 2.5 Omni 3B. Most of these , including Olmo 2 1B , can easy run on a innovative laptop or even a peregrine equipment .
Ai2 says Olmo 2 1B was trained on a data circle of 4 trillion tokens from publically usable , AI - mother , and manually created sources . Tokens are the raw bits of data that mold ingest and generate , with a million token tantamount to about 750,000 words .
On a benchmark appraise arithmetical reasoning , GSM8 K , Olmo 2 1B scores better than Google ’s Gemma 3 1B , Meta ’s Llama 3.2 1B , and Alibaba ’s Qwen 2.5 1.5B. Olmo 2 1B also eclipses the functioning of those three models on TruthfulQA , a test for evaluating actual truth .
Join us at TechCrunch Sessions: AI
Exhibit at TechCrunch Sessions: AI
This exemplar was pretrained on 4 T tokens of gamy - quality datum , following the same monetary standard pretraining into gamey - lineament tempering of our 7 , 13 , & 32B models . We upload intermediate checkpoint from every 1000 steps in education . get at the base model : https://t.co / xofyWJmo85pic.twitter.com/7uSJ6sYMdL
Ai2 has warned that Olmo 2 1B carry risks , however . Like all AI models , it can create “ tough yield , ” include harmful and “ sensitive ” content , the constitution state , as well as factually inaccurate statements . For these reasons , Ai2 recommend against deploying Olmo 2 1B in commercial-grade stage setting .
