Codex + MLX-LM-LoRA MCP: auto-pretrain a tiny LLM on Apple Silicon with a 5-step setup
alexcovo_eth · x · 2026-10-24
ActuallyIsaak shows a reproducible workflow where Codex autonomously pre-trains and iterates on a tiny LLM on Apple Silicon via MLX-LM-LoRA's MCP server and skill. Setup: pip install "mlx-lm-lora[mcp]", add the skill via npx skills add, locate the MCP binary with which mlx-lm-lora-mcp, register it with codex mcp add, then just chat the task. The demo prompt constrains the agent to that stack only, using tiny-random/qwen3, the mlx-community/recyclingtheweb-1k dataset, 100 pre-training steps and global batch size 8 — turning small-model training into an executable agent skill chain.
More from coding & agent
- Dev offers to pay extra for a 'terse mode' that caps agent replies at 5 lines — weswinder · 2026-10-25
- Agent builds 100%-reliable Node.js brotli exploit in 5 unsupervised hours — jcran · 2026-10-25
- Dev builds interactive 3D energy textbook covering full GCSE topic with Opus 5.5 and Three.js — Afinetheorem · 2026-10-25
- Codex CLI 0.164.0 Adds Presets: Save Prompts and Settings, Recall via Ctrl+X P — pvncher · 2026-10-25
- 305M-Parameter Sparse MoE Runs Computer-Use Agent Fully Offline on a MacBook Air — lmoroney · 2026-10-25
- Kling 4 nails one-shot real estate videos with full reference-image consistency — venturetwins · 2026-10-25