Ai2

4B vision-language model that forecasts 3D point trajectories from a single RGB observation frame, 2D query points, and a natural-language action instruction. Built on Molmo 2, it predicts where surface points on object…

• 🧊 3D
• 🎥 Video analysis
• 🤖 Robotics
• 🏃 Motion capture New

Ai2’s open robotics foundation model for real-world robot manipulation. It is an Action Reasoning Model that reasons about 3D structure before acting, supports bimanual manipulation out of the box, and is designed for z…

• 🤖 Robotics
• 🧩 Spatial intelligence
• 👁️ Computer vision assistance

F_pre_8kv_8k_14k_SWA is an AllenAI OlmPool architectural-variation model for studying long-context extension. It is a 7B to 8B class language model trained to 150B tokens total, including 140B pretraining and 10B contex…

• 📚 Large Language Models
• 🔄 Language model optimization

Ai2’s open model for monocular 3D object detection from a single RGB image. It predicts full 3D bounding boxes, including position, size, and orientation in metric coordinates, and supports flexible prompts such as text…

• 🧩 Spatial intelligence
• 🔍 Image recognition
• 👁️ Computer vision assistance 3d

A 32B Qwen 3 based coding agent variant trained on GLM-4.5-Air trajectories, delivering about 46.6% SWE-bench Verified performance as a slightly smaller-teacher alternative to SERA-32B.

• 💻 Coding
• 🔎 Problem solving 🇯🇵 Japanese lessons

Ai2’s flagship 32B Open Coding Agent, a Qwen 3-32B based repository-level coding model trained with Soft Verified Generation that reaches about 49.5% on SWE-bench Verified, rivaling much larger proprietary systems.

• 💻 Coding
• 📚 Code explanations
• 🔍 Code reviews
• 💡 Coding help

An 8B Qwen 3 based coding agent variant trained from GLM-4.5-Air generated trajectories, matching SERA-8B’s 31.7% SWE-bench Verified score while providing an alternative teacher setup for research.

• 💻 Coding

Ai2’s 8B-parameter open-source coding agent, fine-tuned from Qwen 3-8B with Soft Verified Generation to solve repository-level bugs, reaching about 31.7% resolve rate on SWE-bench Verified at 32K context.

• 💻 Coding
• 📚 Code explanations
• 🔍 Code reviews

AI2’s open 4B/7B/8B multimodal model for images and video, delivering state-of-the-art grounded video QA, pointing and tracking that return coordinates and timestamps for events instead of text-only answers.

• 🎥 Videos
• 📝 Writing
• 📷 Images
• 🔍 Multimodal search

Ai2's fully open model family in 7B and 32B sizes, shipped as Base, Instruct, Think and RL Zero variants with a 65K context window. Ai2 releases the whole model flow rather than just weights: the Dolma 3 corpus, every t…

• 💬 Chatting
• 💻 Coding
• 🤔 Logical reasoning
• 🔢 Math

Ai2/AllenAI’s foundation-model collection for remote-sensing tasks using satellite and Earth-observation imagery.

• 🌐 Geospatial analysis
• 🖼️ Image segmentation
• 🌍 Environmental analysis
• 📊 Climate data analysis

AllenAI’s open-source document recognition pipeline and model family that converts PDFs and images into clean text, preserving reading order, tables, equations, and handwriting.

• 🏭 Manufacturing
• 🖼️ Image to text