LLMForge

LLMForge is a native macOS app for working through an LLM workflow from model download to on-device deployment in one window. It is described as private, local, and cloud-free, and the page frames it as a way to avoid stitching together terminal tools and config files.

Its built-in workflow covers browsing and downloading models from HuggingFace, with architecture, size, and RAM requirement shown before download. It also supports dataset curation through CSV or JSONL import, manual labeling, and local AI-assisted pair generation, with outputs in Alpaca or ChatML format and ready for MLX. For training, it runs natively on Apple Silicon using MLX and includes LoRA and QLoRA, live loss monitoring, checkpoints, overfitting detection, persisted sessions, and crash recovery.

LLMForge also offers quantization and export to GGUF or CoreML, side-by-side model comparison, and the ability to save useful outputs back into a dataset for later training runs. A local API server exposes fine-tuned models through an OpenAI-compatible endpoint with SSE streaming, stop-generation support, and live metrics for tokens per second, latency, and token counts. Fine-tuned adapters can be pushed directly to HuggingFace repositories, public or private, from inside the app.

The product is a native Mac application built for developers who ship models. It is fully offline, uses MLX for training and Metal for inference, and is described as built specifically for Apple Silicon, with Apple Silicon, macOS 26+, and 8 GB RAM listed as requirements. The page also says it is free, under 200 MB, and requires no account.