Soup CLI Launches Local 8B Fine-Tuning

Soup CLI makes local fine-tuning meaningfully more accessible by attacking the biggest practical barrier: GPU memory. Its strongest differentiator is the correctness-first layer-streaming implementation, though the beta approach trades speed for hardware flexibility.

• –Streams frozen decoder layers from system RAM instead of loading the full base model into VRAM
• –Supports common PEFT workflows, quantization, automatic batch sizing, and multiple training objectives
• –Covers the full path from dataset inspection and training to evaluation, GGUF export, serving, and Hugging Face publishing
• –Competes with tools such as LLaMA-Factory and Unsloth by prioritizing a unified CLI and configuration-driven workflow
• –Developers should validate training speed and compatibility on their hardware, since layer streaming is slower and remains beta