Codex is Open Sourcing AI models

Source: HuggingFace Date: 2025-12-11 URL: https://huggingface.co/blog/hf-skills-training-codex

Summary

Integration tutorial demonstrating OpenAI Codex automating end-to-end ML fine-tuning workflows via HuggingFace Skills: a single natural language instruction triggers dataset validation, hardware selection, training script generation, job submission to HF, and evaluation. Supports SFT, DPO, GRPO, and GGUF quantization on 0.5B-7B models. Cost: $1-2 (tiny) to $15-40 (3-7B with LoRA). Example: Qwen3-0.6B on codeforces-cots improves HumanEval pass@1 from 0.306 to 0.342.

Implications

Thread: HF as open-source ML hub / agentic patterns. HF Skills turning the HF training platform into a tool callable by AI coding agents is an important distribution move — Codex users can now trigger full fine-tuning runs without understanding HF infrastructure. The “natural language → training job” abstraction is exactly what non-ML-specialist teams want for domain model customization. Requires HF Pro/Team/Enterprise and Codex access, so the user base is gated. Watch whether this pattern (AI agent → HF training job → evaluation loop) gets extended to more agentic frameworks beyond Codex.