Ai2

Today we’re bringing new NSF OMAI compute online with NVIDIA Blackwell Ultra-powered systems, turning a $152M national investment from @NSF & @NVIDIA into a foundation for truly open AI research. 🧵

Really amazing results analyzing what's creative/novel vs. what's copied from Internet data, enabled by the amazing @liujc1998's Infini-gram! infini-gram.io This is also enabled in @allen_ai's OlmoTrace allenai.org/blog/olmotrace where anyone can find matching n-grams between LLM-generated text and its training data.

📢 Save the date!  📌 AI Workflows for Food Systems Research: A Demonstration of AutoDiscovery with the Ai2 Asta Team. 📆 June 4, 9:30 – 10:30 am EDT 💬 Ruben Lozano Aguilera & Bodhisattwa Majumder (@allen_ai); Leroy Mwanzia (@Cipotato); @EliotJonesGarc1 (IFPRI). 🎫 ifpri.zoom.us/webinar/regist… @CGIAR

Very cool to see these conversations happening! This is what openness enables. The "tool that allows you to trace those n-grams directly to their source," is infinigram, AKA OlmoTrace from @allen_ai, created by @liujc1998. x.com/alexolegimas/s…

We're releasing a dataset of 14K HuggingFace models, datasets, papers, & codebases linked by 51K evaluations, fine-tunings, & references, plus the ArtifactLinker code. We hope it helps others find SOTA eval results. 💻 Code: github.com/allenai/artifa… 📊 Data: huggingface.co/datasets/lwaek…

@huggingface Using ArtifactLinker, we found cases where a strong model had never been evaluated on a benchmark it would set – or near-match – the SOTA on. We also found that newer LLMs like Gemma often lose to older DeBERTa models on natural language inference tasks.

Most models are only evaluated on a fraction of the benchmarks out there. ArtifactLinker, our new system, predicts which ones would set a new state-of-the-art on benchmarks hosted on @HuggingFace, then runs the evaluation to verify. 🧵

We release open models so they're available to builders working on problems that matter to them. On Global Accessibility Awareness Day, PointCheck is a fitting example. Read more ↓ allenai.org/blog/global-ac…

PointCheck tabs through a page like a keyboard user and screenshots each step. Molmo locates the focused element; MolmoWeb and Olmo 3 handle the rest. Brendan chose open models for his experimental side project so teams can self-host—no files leave the environment.

Brendan Works is a product manager focused on paratransit services in Seattle. See how he built PointCheck, a website accessibility checker powered by our open Molmo, MolmoWeb, & Olmo 3 models. 👇

Available now in the same sizes as v1: Nano, Tiny, Base. Open weights, open training code. If you're running v1 and v1.1 works for your task, expect significant speedups during fine-tuning & inference. 🤗 Models: huggingface.co/collections/al… 🔗 Blog: allenai.org/blog/olmoearth…

A useful property for researchers: we held the pretraining dataset constant from v1. The differences cleanly isolate the methodological change, not the data or the architecture family.

Today we’re releasing OlmoEarth v1.1. It’s 3x cheaper to run than v1 while delivering the same state-of-the-art performance—and fully open. 🧵

@GoogleResearch @nvidia We’re releasing the first-phase AIMIP dataset + our analysis of it. We hope to continue AIMIP with future phases that expand its scope & scale. 📘 Learn more in our blog: allenai.org/blog/AIMIP 📊 Paper: arxiv.org/abs/2605.06944 🗂️ Dataset: #data" target="_blank" rel="nofollow noopener">github.com/ai2cm/AIMIP/tr…

@GoogleResearch @nvidia We also tested the models on harder scenarios, such as a rapidly warming ocean that was unfamiliar from training In those tests, the models diverged much more—showing that generalization remains a major challenge.

Announcing the AI Model Intercomparison Project (AIMIP), a community effort to support open evaluation of AI climate models. 🌎 It brings together a shared benchmark experiment & dataset to make it easier to compare models side by side over multi-decade simulations. 🧵