Sources
Loading...
Additional media
Loading...
DeepNewz, mobile.
People-sourced. AI-powered. Unbiased News.
First fully open model to beat GPT-4o mini? ✅ Same peft of Qwen 2.5 for 1/3 of training cost? ✅ Everything (data, checkpoints, code, eval) open? ✅ OLMo 2 32B is our best model yet—we scaled the up OLMo 2 recipe and combined it with latest Tülu 3.1 goodies ✨ https://t.co/2bm3yLhVIj
Congrats OLMo team!❤️🔥 (an anecdote from an ex-Ai2er:) in early OLMo meetings 2+ years ago, there was a feeling of hopelessness. it was not clear that open models could ever match closed again. What a change! Thanks for the work, the models, the code, the data, and the hope! https://t.co/XGi70ItKcx
Yo! @allen_ai COOKED - OLMo 2 32B - beats GPT3.5, 4o mini and FULLY OPEN SOURCE 🔥 Model weights on the hub and works with transformers! 🤗 Kudos to the team! https://t.co/nyHNeDVFUj
The open-source AI community has marked a milestone with the release of OLMo 2 32B, a new model that surpasses the capabilities of GPT-3.5 and GPT-4o mini on various multi-skill benchmarks. This model, developed by the OLMo team, is noted for being fully open-source and achieving performance comparable to leading open-weight models like Qwen and Mistral, while requiring only a fraction of the training compute. The announcement highlights the model's availability of data, weights, and code, reinforcing the commitment to open-source principles. The development has been met with enthusiasm from the AI community, reflecting a significant shift in the perception of open models' competitiveness against closed counterparts.
