Self-Training Agents: Hermes Agent, HF Traces, Skills, MCP & Finetuning — Merve Noyan, Hugging Face

Open-source models have caught up. GLM 5.1 is leading the Artificial Analysis intelligence index over closed models, and the gap is closing fast with each release cycle. The practical upside beyond benchmarks: full weight access means you can quantize, fine-tune, and deploy to edge devices or browsers without data leaving your infrastructure. ‪@MerveNoyan‬ walks through the Hugging Face ecosystem built around this: inference providers that route to the fastest or cheapest option per model, benchmark datasets for filtering by SWE-bench or AIME scores directly on Hub, a traces repository type for storing and exploring agent sessions, and skills that plug into coding agents. The closer is a live demo where she asks Claude Code to fine-tune a vision-language model on a dataset by name. The agent calculates VRAM requirements, selects an instance, and kicks off the job. What used to be a day of napkin math is now a prompt. Speaker info: https://x.com/mervenoyann   / merve-noyan-28b1a113a   https://github.com/merveenoyan Timestamps 0:00 Introduction to Open Agent Ecosystem 0:39 Importance of Open Source in Machine Learning 2:36 Hugging Face Hub overview 3:06 Agentic models and Vision-LMs 4:24 Benchmark datasets and model filtering 5:16 Inference providers and model routing 6:50 Local coding agents and tools 7:46 Hermes agents for memory management 9:20 Traces repository for agent sessions 10:22 Tips for finding and serving local models 12:07 Supercharging agents with Hugging Face skills 13:41 Live demonstration of agent-driven fine-tuning 14:41 Training vision models (object detection/segmentation) 15:00 Using Model Context Protocol (MCP) for agents 16:30 Case study: OCR processing for AI papers