Claude can now fine-tune open-source LLMs directly on cloud GPUs

Anthropic’s Claude just got a practical new ability that sidesteps the usual notebook grind: it can now fine-tune open-source language models directly on Hugging Face’s cloud GPUs using a new tool called Hugging Face Skills. This isn’t about generating a Python script and leaving you to wrestle with CUDA drivers. As detailed in a new tutorial, the hf-llm-trainer skill packages the domain knowledge Claude needs—like selecting between LoRA and full fine-tuning, or picking the right GPU for a 0.6B versus a 70B parameter model—to submit jobs, monitor them, and push the finished model to the Hub.

You can give it a command as simple as “Fine-tune Qwen3-0.6B on the dataset open-r1/codeforces-cots” and Claude gets to work. It validates the data, selects the cheapest adequate hardware (a t4-small at roughly $0.75 an hour in this case), and submits the job after showing you a cost estimate—about 20 minutes and $0.30 for that demo. The system then reports back a job ID for monitoring on Hugging Face’s platform while you do other things. This workflow also applies to vision language models with a similarly straightforward prompt.

Under the hood, the skill isn’t a toy. It supports production-grade methods like supervised fine-tuning (SFT), direct preference optimization (DPO), and reinforcement learning with verifiable rewards (RLVR). You can handle models up to 70B parameters, convert them to GGUF format for local deployment, or chain multiple techniques in a pipeline. The setup requires a paid Hugging Face plan and a write-access token, but the integration works beyond Claude Code—it’s compatible with OpenAI Codex and Google’s Gemini CLI, with support planned for editors like Cursor and Windsurf.

The real story here is the shift in who gets to push a training job. This combination of a coding agent and a structured skill dramatically lowers the barrier for fine-tuning, turning what was once a multi-hour, error-prone setup into a natural-language request. If it works reliably, the main constraint stops being technical know-how and becomes simply the cost of cloud compute and your willingness to experiment.