Daily News · 3 min read

Hugging Face AI Updates: June 18, 2026

hugging-face open-source agents robotics models

Hugging Face. The company announced Agentic Resource Discovery (ARD), an open specification that lets agents dynamically search for and discover tools, skills, and other agents at runtime rather than requiring manual pre-installation. ARD defines two components: a static manifest format called ai-catalog.json for publishers to catalog capabilities, and a dynamic registry API that supports natural language search. Hugging Face implemented the spec through its Discover Tool, which indexes thousands of Skills, ML applications, and MCP Servers and is queryable via CLI, REST API, or an MCP endpoint, addressing scalability limits in approaches that hardcode tool configs or dump every tool description into an LLM context window. Source

2. Z.AI Releases GLM-5.2 Open Model on the Hub for Long-Horizon Coding

Hugging Face. Z.AI published GLM-5.2, an MIT-licensed flagship language model built for extended reasoning and coding, with weights publicly available on the Hugging Face Hub at zai-org/GLM-5.2 and support for inference frameworks including transformers and vLLM. The model offers a stable 1-million-token context window, adjustable thinking effort levels, and an IndexShare architecture that the team says reduces computational overhead by 2.9 times at 1M context. On benchmarks it reaches 81.0 on Terminal-Bench 2.1 and 62.1 on SWE-bench Pro, and ranks second only to Claude Opus 4.8 across the FrontierSWE, PostTrainBench, and SWE-Marathon long-horizon engineering benchmarks. Source

3. AI2 Releases MolmoMotion Models and 1.16M-Video Dataset for 3D Motion Forecasting

Hugging Face. The Allen Institute for AI (AI2) released MolmoMotion, a language-guided 3D motion forecasting model that predicts where 3D points on objects will move given video frames, marked points, and written action instructions, built on the Molmo 2 backbone in autoregressive (MolmoMotion-AR) and flow-matching (MolmoMotion-FM) variants. The release ships fully on the Hugging Face Hub with model weights, code, the MolmoMotion-1M dataset of 1.16 million videos with 3D point trajectories and action descriptions, and PointMotionBench, a human-validated benchmark of 2.7K video clips. AI2 reports the model outperforms existing 3D motion forecasting methods and generalizes across object categories without category-specific templates, with applications in robotics planning and video generation. Source

4. AWS Releases Strands Robots SDK Linking the Hub to Robot Hardware via LeRobot

Hugging Face. AWS released Strands Robots, an open-source SDK that integrates Hugging Face’s LeRobot with agent-based orchestration to move from the Hub to physical robot hardware through a unified workflow. Datasets recorded in MuJoCo simulation use the same LeRobotDataset structure as hardware recordings, and the agent code is identical across simulation and real modes, with developers switching by changing a single mode=“real” parameter. The SDK packages recording, training, testing, deployment, and fleet coordination as composable AgentTools within one agent loop and supports multi-robot fleets via Zenoh mesh networking, treating sim-to-real as a deployment detail rather than an architectural problem. Source