| Oct 01, 2025 | Presenting three works at COLM 2025!🎉 Fluid Language Model Benchmarking on language model evaluation, LLMs as Research Tools results from a large scale survey, and our latest recipe for training open language models 2 OLMo 2 Furious. We’re handing out Blu-Ray discs with our OLMo 2 model weights and hosting an Ai2 event! Ready for some bagels 🥯. Also, I have a shiny new backpack to show off… |
| Jun 01, 2025 | Molmo and PixMo received a Best Paper Honorable Mention at CVPR 2025!🏆 Congrats to Matt, Chris and our Ai2 team! |
| May 15, 2025 | Excited to welcome our 2025 interns: Mayee Chen, Yapei Chang, Amanda Bertsch, and Alexis Ross! 🎉 |
| May 01, 2025 | Organize the Web accepted to ICML 2025! Data isn’t just about “quality”, slice and dice by “topic” and “format” domains. Congrats Alex & see everyone in Vancouver 🇨🇦! |
| May 01, 2025 | DrawEduMath won an Outstanding Paper Award at NAACL 2025! 🏆 Congrats to Lucy and the team! |
| Mar 13, 2025 | We released our largest and best model yet! OLMo 2 32B is trained using the same recipe from 2 OLMo 2 Furious, comparable base model performance to some of the best open-weight models like Qwen and Gemma. After instruction tuning, it’s the best fully open model to reach GPT 3.5/4o mini performance. Our blog post says more. As always, download the model weights, data, and everything on HuggingFace! |
| Jan 15, 2025 | OLMoE accepted as an Oral (top 1.8% of 11.6K submissions) at ICLR 2025! Congrats Niklas. I’m also going to be giving a talk on data curation for OLMo 2 at the Data Problems for Foundation Models workshop. See you all in Singapore 🇸🇬! |
| Dec 01, 2024 | Giving a tutorial on Opening the Language Model Pipeline at NeurIPS 2024 with my colleagues Akshita Bhagia and Nathan Lambert! We’ll cover data preparation, model training, and adaptation methods using open software and data. Excited to share tips, tricks, and otherwise inaccessible details from building OLMo! |
| Nov 26, 2024 | OLMo 2 is here! In our latest paper 🚗 2 OLMo 2 Furious 🔥, we discuss everything we’ve learned since OLMo 1 with deep dives into 🚖 stable pretraining and 🚔 mid-training which uses learning rate annealing, data curricula, and model checkpoint averaging. Our training recipe is state-of-the-art with respect to training FLOPs to performance!📈 Check out the blog post and download our 7B and 13B model weights, data, etc on HuggingFace! |
| Oct 01, 2024 | Excited that our Semantic Reader paper is published in Communications of the ACM! 🥳 This paper synthesizes our five years of AI and HCI research (50 researchers, 12 institutions) aimed at understanding reading challenges faced by scholars and how AI-powered intelligent interfaces can help. Check out the paper here! |
| Sep 25, 2024 | Molmo is out! Molmo is our family of open, late-fusion image 👀 + text 💬 language models trained using a really high-quality dataset of images + dense captions / task demonstrations! ✅ Read the paper here, ✅ play with the model here, ✅ download the weights here, and ✅ look forward to our dataset release soon! |
| Sep 03, 2024 | OLMoE is out! Our first mixture of experts model in the OLMo family 🎉 OLMoE has only 1B active params but matches perf of larger dense models 🫨 and comes released with: ✅ weights ✅ data ✅ code ✅ ckpts ✅ logs ✅ detailed paper! Download the weights here and read the paper here! |
| Aug 14, 2024 | So proud to see both our OLMo and Dolma papers win 🏆 Best Paper awards 🏆 at ACL 2024 🇹🇭 |
| Jul 25, 2024 | Excited to be speaking at Gen Law workshop at ICML 2024 in 🇦🇹! I’ll be sharing fun pretraining data curation stories from OLMo, and my slides have cats! 🐈 |
| Jun 06, 2024 | |