‘Tis the week for small AI fashions, it appears.
On Thursday, Ai2, the nonprofit AI analysis institute, launched Olmo 2 1B, a 1-billion-parameter mannequin that Ai2 claims beats similarly-sized fashions from Google, Meta, and Alibaba on a number of benchmarks. Parameters, generally known as weights, are the inner elements of a mannequin that information its conduct.
Olmo 2 1B is accessible below a permissive Apache 2.0 license on the AI dev platform Hugging Face. Unlike most fashions, Olmo 2 1B could be replicated from scratch; Ai2 has supplied the code and knowledge units (Olmo-mix-1124, Dolmino-mix-1124) used to develop it.
Small fashions won’t be as succesful as their behemoth counterparts, however importantly, they don’t require beefy {hardware} to run. That makes them way more accessible for builders and hobbyists contending with the restrictions of lower-end and client machines.
There’s been a raft of small mannequin launches over the previous few days, from Microsoft’s Phi 4 reasoning household to Qwen’s 2.5 Omni 3B. Most of those — and Olmo 2 1B — can simply run on a contemporary laptop computer or perhaps a cell gadget.
Ai2 says that Olmo 2 1B was educated on a knowledge set of 4 trillion tokens from publicly obtainable, AI-generated, and manually created sources. Tokens are the uncooked bits of information fashions ingest and generate — 1 million tokens is equal to about 750,000 phrases.
On a benchmark measuring arithmetic reasoning, GSM8K, Olmo 2 1B scores higher than Google’s Gemma 3 1B, Meta’s Llama 3.2 1B, and Alibaba’s Qwen 2.5 1.5B. Olmo 2 1B additionally eclipses the efficiency of these three fashions on TruthfulQA, a take a look at for evaluating factual accuracy.
Techcrunch occasion
Berkeley, CA
|
June 5
BOOK NOW
Ai2 warns that that Olmo 2 1B carries dangers, nevertheless. Like all AI fashions, it could produce “problematic outputs” together with dangerous and “delicate” content material, the group says, in addition to factually inaccurate statements. For these causes, Ai2 recommends in opposition to deploying Olmo 2 1B in business settings.
