Skip to main content ->
Ai2

Latest research

April 30, 2026

AstaBench update: New results, plus adoption from industry

AstaBench’s latest update adds new frontier-model results, including GPT-5.5, and highlights growing adoption from groups including the UK AISI, General Reasoning, Elicit, SciSpace, Distyl AI, and EvoScientist.
Read post
April 23, 2026

Introducing OlmoEarth embeddings: Custom embedding exports from OlmoEarth Studio for downstream analysis

OlmoEarth Studio now lets users export custom Earth-observation embeddings from our OlmoEarth foundation models and use them for tasks like similarity search, few-shot mapping, change detection, and unsupervised exploration.
Read post
April 23, 2026

OlmPool: How small architectural choices compound to undermine long context extension

OlmPool is a controlled suite of 26 models showing how small architecture choices can compound to make long-context extension much harder, even when training data and extension recipes are held constant.
Read post
April 20, 2026

Train separately, merge together: Modular post-training with mixture-of-experts

BAR is a recipe for post-training language models one capability at a time—train domain experts independently, merge them into a single mixture-of-experts model, and upgrade any expert without impacting the others.
Read post
April 13, 2026

Evaluating agents for scientific discovery

Two benchmarks developed at Ai2 – ScienceWorld and DiscoveryWorld – reveal that even incredibly strong AI science agents struggle with problems human scientists solve routinely.
Read post
April 7, 2026

Introducing WildDet3D: Open-world 3D detection from a single image

WildDet3D is an open model that predicts 3D bounding boxes from a single image. It generalizes across cameras and object categories, and folds in depth signals when available—alongside a new dataset of verified 3D annotations.
Read post
March 24, 2026

MolmoWeb: An open agent for automating web tasks

Introducing MolmoWeb, an open visual web agent that navigates and completes tasks in a browser using screenshots alone, along with MolmoWebMix, the largest public dataset for training web agents.
Read post
March 18, 2026

MolmoPoint: Better pointing architecture for vision-language models

MolmoPoint is a new vision-language model architecture that replaces text-based coordinate outputs with a more natural, token-based pointing mechanism that directly selects regions from visual features.
Read post
March 11, 2026

MolmoBot: Training robot manipulation entirely in simulation

MolmoBot is an open robotic manipulation model suite trained entirely in simulation—demonstrating zero-shot transfer to real-world robots without any real-world data collection or fine-tuning.
Read post
1-9Next