AI
Open Models Now Match Closed Models
Open models have caught up on benchmarks. GLM 4.1 currently tops SWE-bench. The argument that open-source quality lags is finished. With weights in hand, you can quantize Gemma 4 to 4-bit and run it on a 24 GB L4 GPU. You can shrink, fine-tune, or push to edge devices and browsers. When a closed API silently degrades, you have no recourse. With open weights, nothing changes without you knowing, and user data never leaves the device.
Local Coding Agents and Hermes Agents
For local coding agents, Noyan’s top picks are Py (backed by llama.cpp) and llama-agent, a binary baked into llama.cpp that starts a model by HF Hub ID. Hermes agents go further: memory management, Slack and WhatsApp integration, set up by a wizard that just asks for your keys. She tested it with the Hermes agent fixing Slack integration (08:29), where a broken Slack connection was repaired autonomously by GLM 4.1. Her recommendation: run Hermes with GLM 4.1 or via HF inference providers if you want a capable open model with minimal setup.
Agent Traces as a First-Class Dataset Type
HF Hub now has a dataset repository type called traces. Push your Codex, Claude Code, or Py session logs to a Hub repo, open the dataset viewer, click the traces column, and you get a parsed, browsable view of every agent step. The data stays yours, and if you accumulate enough of it, you can fine-tune a model on those traces directly from Hub. Hermes agent support for traces is coming soon. Two lines of path config is all the upload requires.
Skills That Let Agents Train and Deploy Models
HF Skills wire coding agents into Hub infrastructure. The LLM Trainer skill accepts a plain prompt, calculates VRAM, picks an instance, and launches the job. Noyan asked Claude Code to train Qwen2-VL on LLaVA-Instruct-Mix; the agent handled batch-size math without manual config. The same pattern OCR’d 30,000 Hub research papers: ask an agent to write the script, pick a model from OCR-bench, compute costs, and run it. Noyan’s colleague did all of that via MCP spaces image generation demo (15:43), through prompting alone.
Notable Quotes
it. And it’s absolute guaranteed privacy Merve Noyan · ▶ 1:43
just trains which to me is like a sci-fi Merve Noyan · ▶ 6:23
stuff. So we OCR 30 30,000 papers uh Merve Noyan · ▶ 16:55
Key Takeaways
- Open models like GLM 4.1 now top SWE-bench, matching closed models on coding benchmarks.
- Hermes agents handle memory and tool use locally or via HF inference, with minimal setup.
- HF Skills let agents trigger fine-tuning jobs, build demos, and query datasets without manual infrastructure work.