‘Tis the week for small AI models, it seems.
On Thursday, Ai2, the nonprofit AI research institute, released Olmo 2 1B, a 1-billion-parameter model that Ai2 claims beats similarly-sized models from Google, Meta, and Alibaba on several benchmarks. Parameters, sometimes referred to as weights, are the internal components of a model that guide its behavior.
Olmo 2 1B is available under a permissive Apache 2.0 license on the AI dev platform Hugging Face. Unlike most models, Olmo 2 1B can be replicated from scratch; Ai2 has provided the code and data sets (Olmo-mix-1124, Dolmino-mix-1124) used to develop it.
Small models might not be as capable as their behemoth counterparts, but importantly, they don’t require beefy hardware to run. That makes them much more accessible for developers and hobbyists contending with the limitations of lower-end and consumer machines.
There’s been a raft of small model launches over the past few days, from Microsoft’s Phi 4 reasoning family to Qwen’s 2.5 Omni 3B. Most of these — and Olmo 2 1B — can easily run on a modern laptop or even a mobile device.
Ai2 says that Olmo 2 1B was trained on a data set of 4 trillion tokens from publicly available, AI-generated, and manually created sources. Tokens are the raw bits of data models ingest and generate — 1 million tokens is equivalent to about 750,000 words.
On a benchmark measuring arithmetic reasoning, GSM8K, 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 performance of those three models on TruthfulQA, a test for evaluating factual accuracy.
Techcrunch event
Exhibit at TechCrunch Sessions: AI
Secure your spot at TC Sessions: AI and show 1,200+ decision-makers what you’ve built — without the big spend. Available through May 9 or while tables last.
Exhibit at TechCrunch Sessions: AI
Secure your spot at TC Sessions: AI and show 1,200+ decision-makers what you’ve built — without the big spend. Available through May 9 or while tables last.
Berkeley, CA
|
June 5
BOOK NOW
This model was pretrained on 4T tokens of high-quality data, following the same standard pretraining into high-quality annealing of our 7, 13, & 32B models. We upload intermediate checkpoints from every 1000 steps in training.Access the base model: https://t.co/xofyWJmo85 pic.twitter.com/7uSJ6sYMdL— Ai2 (@allen_ai) May 1, 2025
Ai2 warns that that Olmo 2 1B carries risks, however. Like all AI models, it can produce “problematic outputs” including harmful and “sensitive” content, the organization says, as well as factually inaccurate statements. For these reasons, Ai2 recommends against deploying Olmo 2 1B in commercial settings.
Topics
AI, AI2, open source
Kyle Wiggers
AI Editor
Kyle Wiggers is TechCrunch’s AI Editor. His writing has appeared in VentureBeat and Digital Trends, as well as a range of gadget blogs including Android Police, Android Authority, Droid-Life, and XDA-Developers. He lives in Manhattan with his partner, a music therapist.
View Bio
May 13, 2025
London, England
Get inside access to Europe’s top investment minds — with leaders from Monzo, Accel, Paladin Group, and more — plus top-tier networking at StrictlyVC London.
REGISTER NOW
Most Popular
Microsoft CEO says up to 30% of the company’s code was written by AI
Maxwell Zeff
Google launches AI tools for practicing languages through personalized lessons
Aisha Malik
Indian court orders blocking of Proton Mail
Jagmeet Singh
Hugging Face releases a 3D-printed robotic arm starting at $100
Kyle Wiggers
Bezos-backed Slate Auto debuts analog EV pickup truck that is decidedly anti-Tesla
Sean O'Kane
Wait, how did a decentralized service like Bluesky go down?
Sarah Perez
Perplexity CEO says its browser will track everything users do online to sell ‘hyper personalized’ ads
Julie Bort
Fintech Bench conducts layoff while others still work month-to-month do sex
FutureHouse releases AI tools it claims can accelerate science do sex