Descriptions:
Merve Noyan from the Hugging Face open-source team delivers a broad survey of the current open-model landscape alongside several first-party announcements of new Hugging Face Hub features targeting agentic AI workflows. The talk opens by reframing the open-vs-closed debate in practical terms: with the Artificial Analysis intelligence index showing open models like GLM 5.1 at or near the top of composite benchmarks, the performance gap has largely closed—and open weights offer additional guarantees around privacy, edge deployment, and auditability that closed APIs cannot match.
Several new Hub capabilities are demonstrated live. A “traces” dataset repository type now stores agent session logs from tools including Claude Code and Codex, with a structured viewer for exploring individual traces and a path to fine-tuning models on captured sessions. A benchmark datasets filter surfaces popular evaluations—SWE-bench Pro, AIME, Humanity’s Last Exam—with open models ranked by score, making model selection for specific use cases more tractable across the Hub’s roughly three million hosted models. An MCP server lets any LLM workflow plug directly into the Hub, while inference providers handle routing across Groq, Cerebras, and others with per-model tool-use support visible in the UI.
Noyan also highlights Hermes agents as a recommended framework for running persistent, memory-managed local agents on open models, recommending GLM 5.1 as a current top pick and citing a personal anecdote where the model autonomously fixed a broken Slack integration when prompted through a Hermes agent session.
📺 Source: AI Engineer · Published May 13, 2026
🏷️ Format: Keynote Launch
