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

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.

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.

A recap of Ai2's week at NVIDIA GTC 2026, covering panels on open models, live demos of Olmo Hybrid and Asta AutoDiscovery, and conversations on coding agents, hybrid architectures, and robotics.

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.

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.

Introducing MolmoBot and MolmoSpaces, an open foundation for training real-world robots to advance science.

Olmo Hybrid is a fully open 7B language model that combines transformer attention with linear RNN layers to achieve greater expressivity and significantly improved data and compute efficiency compared to pure transformer models.

The Asta Interaction Dataset (AID) contains real researcher queries revealing how scientists actually use AI-powered research tools, and where their habits diverge from what tool builders expect.

PreScience is a new benchmark that evaluates whether AI can forecast how science unfolds end-to-end, from team formation through eventual impact.

Olmix is a framework for language model data mixing that provides empirically grounded defaults and efficient reuse techniques.

AutoDiscovery explores data autonomously, generating its own hypotheses to surface surprising findings that researchers might never have thought to look for.

Learn about how researchers are using AutoDiscovery, our scientific discovery tool, to make transformative impact across their fields.

MolmoSpaces is our new open platform for embodied AI that provides physics-grounded scenes, objects, and grasp annotations to train and evaluate generalist robotic policies.

How2Everything is an open framework for evaluating and improving how well LLMs generate step-by-step procedures.

We're excited to share that our paper “Synthesizing scientific literature with retrieval-augmented language models” has been accepted to Nature.

Theorizer is a system that automatically reads scientific literature and synthesizes structured, testable theories.

SERA is the first in our family of Open Coding Agents, achieving state-of-the-art performance at low cost.

HiRO-ACE is an AI framework that makes kilometer-scale climate simulation dramatically more accessible, generating decades of precipitation data for any region of the globe.

Bolmo is new a byte-level family built by adapting Olmo 3 into a fast, flexible byte-based model with a short extra training run.

NeuroDiscoveryBench is a benchmark to test how well AI systems can answer questions grounded in real-world neuroscience data.

Molmo 2, a new suite of state-of-the-art vision-language models with open weights, training data, and training code, can analyze videos and multiple images at once.

Our new flagship Olmo 3 model family empowers the open source community with not only state-of-the-art open models, but the entire model flow and full traceability back to training data.

We introduce Deep Research Tulu (DR Tulu), an open post-training recipe and framework for long-form deep research agents.

Introducing OlmoEarth Platform, an open, scalable, end-to-end system that turns multi-sensor Earth data into constantly-updating, decision-ready insights.